In the rapidly evolving field of academic research, efficient collaboration between different disciplines is essential for the progress of scientific innovation. Unified metadata facilitates and empowers this synergy, as it can substantially improve the practicality of research data across various scientific areas.
But in practice, metadata is often fragmented and comes in different formats. It is often scattered across multiple databases and systems like PyRAT, RSpace, Zotero, and other research platforms.
This makes the integration and analysis of the data a much more tedious process which leads to inefficiencies and research barriers.
This thesis will incorporate both imperative programming methods and AI-driven approaches, such as Large Language Models (LLMs) and Natural Language Processing (NLP) to automate and improve data management and evaluation. Furthermore, the capabilities of NFDI4Health—the National Research Data Infrastructure for Personal Health Data–will be leveraged to integrate and access streamlined metadata from diverse sources within the scientific research domain.
This thesis presents solutions with the help of the described systems and technologies and compares them in order to make research metadata more accessible to scientists, with the aim of optimizing the standardization of metrics from different formats through automation. Establishing an interoperable data infrastructure facilitates collaboration between researchers and lays the groundwork for larger projects.

The aim of this work is to develop an automated solution for detecting cardiac organoids and counting their pulses in microscopic video recordings. Organoids are three-dimensional cell structures that mimic the function of specific organs. In the case of cardiac organoids, analysis of their pulse behaviour can provide valuable information for biomedical research, particularly in the study of heart-like functions. Currently, manual analysis of such videos is time-consuming, error-prone and inefficient. The proposed solution uses classical image processing and signal analysis techniques to automate the detection and counting process. The architecture is based on a modular, object-oriented design that includes several components: User Interface, Backend, Image Segmentation, Signal Processing and Data Management. The user interface allows researchers to upload videos and define regions of interest (ROI). The backend processes the video data by extracting frames, pre-processing images and segmenting cells using algorithms such as Gaussian blur, binary thresholding and watershed segmentation. The signal processing component then tracks the intensity changes of the cells to count the pulsations. The results show that this automated system provides a practical approach to analysing cell pulsations with reduced user effort, low power requirements and increased efficiency, although there are some limitations in the accuracy of detecting weak pulsations. This solution is intended to assist researchers by providing a tool for initial analysis that can be used as a basis for further manual review or as a stepping stone to more advanced methods in future iterations.

Atrial Fibrillation (AF) is one of the most prominent cardiovascular diseases and a leading cause of mortality. An essential diagnostic technique for identifying this cardiac abnormality involves the identification of R-peaks within an Electrocardiogram (ECG). The R-peak is the easiest-to-identify landmark of the ECG and represents the ventricular depolarization of the heart. Detecting this peak enables the inference of heart rhythm irregularities, a central characteristic of AF.
Given the crucial role of R-peak detection in AF diagnosis, software solutions have been developed with algorithms designed to identify this feature. These software are implemented primarily in Python, targetting mainly stationary devices. Recently, there has been a growing trend in portable devices, e.g., AppleWatch, which can record ECGs and are accessible to diverse users. These can be used for diagnosing cardiac anomalies. The automatic evaluation of the user’s recording can be done on mobile phones. Most mobile phones lack a Python interpreter, making them incompatible with existing software. While there are a few R- peak detection libraries in Java designed for Android-based systems, there is a lack of iOS implementation with this functionality.
ECGs recorded by a diverse user using portable devices can be affected by cardiac anomalies and various noise sources that may distort the R-peaks in the ECGs. This necessitates additional features, such as context-aware algorithm selection and signal quality assessment.
The primary objective of this thesis is to close the lack of an R-peak detection software solution and propose an architecture for a software package for cardiac signal analysis. This architecture is implemented in a Swift package named PeakSwift, which incorporates nine R-peak detection algorithms, context-aware selection of R-peak detection algorithms, and two methodologies for assessing signal quality.
The software architecture adheres to industry standards and is designed to be extensible, modular, and scalable. Its open-source nature also makes it a strong foundation for developing additional cardiac analysis features, encouraging contributions from other scientists.
During benchmarking, PeakSwift exhibited remarkable accuracy when tested on an exten- sive dataset of 500 ECGs recorded by diverse subjects in a telehealth setting. Compared to the reference implementation, the R-peak detection algorithms in PeakSwift precisely identified R-peaks in 98.375% of cases. The signal quality assessment correctly classified the signal quality in 98.5% of cases. Regarding runtime performance, PeakSwift proved to be highly competitive and, on average, outperformed the reference implementation by 19.53 ms. When compared to Java solutions, PeakSwift achieved similar runtimes.
In conclusion, this thesis produced a highly precise, runtime-performant, and extensible software package for fundamental cardiac analysis in the iOS environment.

The prevalence of mental health disorders globally is a pressing concern, with many affected individuals not receiving necessary care. Early detection is essential for the prevention of the exacerbation of mental issue symptoms, but a consid- erable treatment gap exists due to numerous challenges. Traditional methods of mental health assessment are often resource-intensive and subject to problems such as privacy concerns and biases, creating a need for innovation solutions. The ubiquitous nature of digital communication, particularly through messaging ap- plications, offers an unprecedented chance to explore the rich source of messaging app data.
This thesis proposes a privacy-preserving approach to recognize communication patterns within messaging app chat data using machine learning techniques. A Swift package and an iOS application are design and developed, which is able to perform a comprehensive on-device analysis, thereby ensuring user privacy and data security. Furthermore, an analysis of aggregated chat data from five partici- pants was conducted, with focus on factors such as demographic profiles, message and emoji frequency, sentiment analysis, temporal analysis, and sentiment-related correlations. The results showed that metrics such as average message length, av- erage reply time, and the percentage of messages sent during different timeframes could potentially serve as indicators for mental health issues. These insights high- light the potential of leveraging chat data for mental health assessment, opening the door for further research in this emerging field. However, given the limited sample size and novelty of this research area, more extensive research on chat data is needed to refine the methodology, generalize the findings, thereby explore the full potential of chat data in mental health assessment.

Depression stands as a significant health challenge, impacting millions of individuals annually, with a considerable number of cases remaining undiagnosed and a high relapse rate [1], [2]. Detecting depression poses challenges, underscoring the necessity for effective and privacy-preserving detection methods. Additionally, depression demonstrates correlations with seemingly unrelated behaviors. Mobility patterns, in particular, have emerged as potential indicators of depression. As highlighted by several studies showcasing a robust correlation between specific location-related features and negative mood changes [3], [4].
This thesis explores the feasibility of utilizing GPS location data from mobile devices for privacy-preserving monitoring of depression-related features. Implemented through the MoodLocator and companion MoodLocator Analysis apps on iOS devices. This study focuses on analyzing location-based data to detect patterns indicative of depression. Two participants tracked for 168 days between June 16 and December 1, revealed distinct mobility and home-stay behaviors. Correlations between traveled distance, variance in locations, and entropy values provided valuable insights.
Our study revealed results that corroborate existing studies, underscoring the sig- nificance of location data in mental health research. However, these results must be interpreted with caution due to our limitations such as a limited sample size. The apps developed during this thesis are presented as reusable tools for future work. Subsequent work can build upon these applications, addressing efficiency concerns, taking advantage of on real-time processing advancements, integrating periodic mood surveys, and developing more robust predictor models. This thesis contributes to the emerging field of on-device depression monitoring, contributing to the broader landscape of mental health research and technology.

Atrial fibrillation (AF) is a serious disease associated with a fivefold increase in stroke risk. If atrial fibrillation is recognized, then the risk of stroke can be reduced through anticoagulation. However, it is often undiagnosed. Recently, devices such as smartwatches with ECG functionality and handheld ECG recorders have been developed. These provide the opportunity to identify and screen for atrial fibrillation in a telehealth setting, particularly when used with automated analysis algorithms to identify ECGs which show signs of atrial fibrillation. Accurate beat detection is essential in automated analysis to identify the irregularly irregular inter-beat intervals associated with atrial fibrillation. It is not yet clear how well ECG beat detection algorithms perform in the telehealth setting, nor which algorithm performs best. Previous work has only compared beat detection algorithms on data from clinical settings or studied a limited number of algorithms. This study evaluated 16 open-source algorithms across seven databases, six of which were publicly available. Performance was assessed on ECGs with sinus rhythm (SR) and atrial fibrillation, ECGs of high and low quality, and on ECGs collected in clinical and telehealth settings. Additionally, a synthetic database was generated to test the impact of atrial fibrillation (AF) and different noise types in controlled conditions. Across the experiments, unsw, which was implemented in MATLAB, continuously got high F1 scores (used as a measure of performance) but was comparatively slow. Especially for noisy data nk, implemented in Python, performed well and was among the fastest algorithms. Algorithm performance significantly decreased from clinical to telehealth settings. In addition, performance on high-quality signals was significantly higher than on low-quality ones. The evaluation of the influence of atrial fibrillation revealed that the algorithms hamilt, nk, and unsw performed better on atrial fibrillation than sinus rhythm. An investigation of performance differences between female and male subjects did not yield conclusive results. A comparison of performance in the presence of different noise types indicated that performance decreased with noise, and that performance was most impacted by motion artifacts, followed by electrode movements, with baseline wander affecting performance the least. nk and unsw should be considered for use in atrial fibrillation screening as they performed well overall, specifically in telehealth settings, and did not show reduced performance on atrial fibrillation data. The algorithms, evaluation code, and synthetic database, are all openly available.

Artificial intelligence (AI), in particular machine learning, has risen to prominence in the recent years. However, conventional machine learning possess risks of severe data breaches because they store their training data in one location. The introduction of strict data protection law, such as GDPR in European Union, further motivates machine learning to be conducted in a decentralized manner. Federated learning enables mobile devices to collaboratively train without sharing their training data. This opens up possibilities for machine learning to train on mobile devices with sensitive data, for instance, health data that are collected by wearables like Apple Watch, without compromising privacy. Recently, federated learning has received increased attention as mobile devices can now deliver on-device training, thanks to technological advances that equip mobile devices with powerful AI chips.
Despite the increased attention in federated learning and advancement on on-device training, there is still little to no support for federated learning in iOS platform. In this thesis, the capabilities of on-device optimization in iOS are explored and a prototype client for iOS, that can participate in a federated learning setting, is presented and evaluated. The purpose of the prototype is to create a stable federated learning system that preserves privacy and enables further development and research in large scale to be conducted in iOS platform. The requirements for the prototype are established using requirements elicitation, aiming to address privacy, scalability and heterogeneous clients challenges in federated learning. The prototype implementation is based on Flower, a federated learning framework with language-, framework-, and platform- agnostic capabilities. This enable the prototype to be scalable and has a plug-and-play characteristics, allowing various components such as privacy-preserving techniques and machine learning frameworks to be integrated easily.

Injuries can be reduced by regular functional movement analysis. Performing a manual evaluation of these is time-consuming and should be automated. Despite the fact that much research is being conducted in the field of automated motion capture and analysis, there is still a lack of systems that can be used remotely by patients. In this thesis, an iOS application is developed, which is capable of tracking motion using the ARKit framework without requiring additional equipment other than the device itself. Data gathered from the motion detection is used to evaluate the Single-Leg Squat Test or the Ankle Lung Test. This system can measure knee and ankle angles, along with measuring knee and hip deviations. Upon completion of the test, the results are visually displayed in the app. This thesis shows that it is possible to implement motion capture and analysis exclusively using a mobile device. Data is tracked reliably. However, the accuracy of the tracked data could not be verified within this thesis.

Insulin treatment in diabetes mellitus requires rigorous self-monitoring to avoid hypo- and hyperglycemia. Related work uses data from continuous glucose monitoring (CGM) devices to predict deviations from the blood glucose norm. However, these devices can impose a medical and financial burden on individuals with diabetes.
This thesis examines the feasibility of commercially-available wearable devices to (1) improve the forecast of glucose values from CGM data and (2) predict dysglycemic events on their own. The analysis relies on the RADAR-dataset of 33 individuals with Type 1, Type 2 or Type 3c diabetes under free-living conditions. In addition to the physiological signals captured by an Empatica and a Garmin smartwatch (e.g., heart rate variability and electrodermal activity), features in the following domains are designed: circadian rhythm, physical activity, sleep and stress. SHAP (SHapley Additive exPlanations) values are applied to assess the feature attributions.
Of all factors, only the circadian rhythm improves the statistical accuracy of the glucose forecasts. It reduces the root-mean-squared-error by 0.86 ± 2.51% (p < 0.01) in the 60-minute and by 2.36 ± 2.53% (p < 0.01) in the 120-minute horizon. Augmenting CGM with smartwatch data currently does not have a significant impact on clinical accuracy. The number of data points in the clinically unacceptable Clarke-Error-Grid zones remains unchanged. The last available glucose value is identified as the most important factor.
The performance of dysglycemic event prediction solely from smartwatch data is currently insufficient for clinical use. The best-performing model combines physiological signals with stress to predict dysglycemia two hours in advance. It achieves an average area under the curve score of 0.72 ± 0.1 and a balanced accuracy of 0.57 ± 0.15. Sustained elevation of electrodermal activity has the strongest correlation with the likelihood of predicting hypoglycemia in two hours.

The increasing life expectancy lead to a higher risk of age-related diseases such as dementia. Still, about 50% of all dementia cases remain undetected. The usual screenings take too long to include in routine doctor visits and are unsuitable to be performed independently. Although many researchers have addressed the problem, the solutions developed remain proactive and are strenuous for those affected.
As global smartphone coverage increases yearly and even most people over 70 use smart- phones regularly, it is natural to assess the possibilities of the most popular function: text messaging.
This thesis proposes an alternative iOS smartphone keyboard. Although it does not differ from the original system keyboard in terms of usage, it continuously collects metrics about typing behavior. The particular focus is on the collection and processing of the data in a privacy-preserving way. The typed data of the individual keys are prepared such that they no longer contain any sensitive data when exported. In general, more emphasis is on metadata, such as the typing speed, than on the actual typed text.
Initial results were collected and analyzed in a small study including 11 subjects. This involved collecting data from healthy subjects and analyzing them for statistical distributions and correlations. The keyboard’s functionality was successfully validated, and the generated data indicate a uniform distribution among the healthy participants. Concluding, the thesis provides an outlook on a more extensive planned study, which aims to get closer to the goal of detecting neurocognitive impairment as occurring in dementia.

In recent years the use of mobile devices has steadily increased. These devices record multitudes of personal data that can be used to create prediction models and generate statistics. To use this emerging data source, one needs a reliable way to collect and access this data, which conforms to the high privacy standards for personal data.
We analyze prior work on secure aggregation and describe a privacy preserving aggregation protocol based on them. We provide a concept implementation of this protocol and develop a framework for mobile applications around it. It enables third parties to extend their mobile applications with the ability to aggregate personal data from their users while preserving the data privacy of each user towards the third party. The implementation consists of an Android application and a containerized server. The framework lowers the barrier of entry for the privacy preserving usage of personal data. The functionality and viability of the framework is demonstrated through a case study about daily step counts.

Diagnosing dementia is a challenging task because the clinical picture is
very diverse and often resembles that of other diseases such as normal aging
or depression. In addition, close monitoring of patients is complicated by
time-consuming and personnel-intensive diagnostic procedures. As a result,
up to 50 percent of dementia cases go undiagnosed. To improve patient monitoring,
smartphones can be used to continuously look for symptoms, rather
than just at follow-up appointments. This thesis proposes a solution to aid
in the early recognition of dementia through the analysis of text input on
smartphones. Specifically, an iPhone application is developed which monitors
the typing behavior of the user through a custom keyboard extension.
Subsequently, the gathered data is analyzed locally on the device to recognize
conspicuous patterns in the users‘ language. The chosen approach is
designed to be used independently of any suspicion or diagnosis to expose
signs of dementia as early as possible. The theoretical foundation of the
analysis is a set of seven measurements that are derived from literature research
in the field of language production of dementia patients: vocabulary
size, text length, repetitions, empty words, fragments, subordinate clauses
and part-of-speech tagging. These have been verified in various studies to
be capable of distinguishing between dementia patients and healthy control
subjects. Based on these criteria, the app employs a technique of sliding windows
to identify trends that align with findings in the literature and assesses
the risk for dementia using three categories: low-, medium- and high-risk.
All results are accessible to the user through the respective iOS app for
inspection. Although the app does identify trends in the typing behavior
correctly, the sensitivity of the analysis is very high because even the slightest
trend is exposed. To evaluate the applicability in the diagnostic process
of dementia and verify the correctness of the results, a separate study must
be conducted as no real-world data was available during the writing of this
thesis. Instead, the texts from a German author who reportedly su↵ered
from dementia were analyzed to provide a use case. Despite limited information
value, the results reveal the potential of this approach by triggering
a medium-risk warning. This method o↵ers unique opportunities for further
research due to its seamless integration into everyday life and can easily be
extended by more measurements.

Eye examination tests have primarily been conducted by healthcare professionals such as optometrists, opticians or ophthalmologists. Conducting standardised eye examination tests on mobile phones could be a new way to perform these tests, allowing for more frequent checks and broader access by the population.
This thesis documents the development process behind the creation of a visual acuity test application. The thesis commences with a brief review of the history of visual acuity testing, and a comparison of the various existing vision testing applications. Then, the development process is dissected, starting with mock-ups, going through the implementation details and continuing with the evaluation of the application. The resulting application displays individual characters on the device’s screen, and asks the user taking the test to pronounce the character displayed.
To gather information about the test performed, sensors such as the TrueDepth camera system, and microphones with the help of speech-to-text machine learning models are used. The data obtained from these sources are assessed in the evaluation section. For this purpose, an additional application was created to specifically support with data collection and model testing, and a script was developed for analyzing the data. Due to the lack of features defining the spelled letters of the alphabet, it proved difficult for the trained machine learning model to differentiate between them. However, pronouncing numbers or directions has proved to be a suitable alternative. In addition, the TrueDepth camera system has proven to be reliable in measuring the distance between the phone and the user’s face. The only limitation is that these sensors have a maximum range of one metre, which restricts the distance at which the test can be performed.
The application resulting from this thesis lays a foundation for the creation of vision impairments detection software, which can be further expanded in the future.

During medical studies, students are faced with learning to perform complex
surgical procedures under a microscope. To mimic the behavior of human
organs, they often use animal models for their practice. Due to the
di culty of supervising and guiding students, an immense amount of mice
are utilized. To reduce the number of mice used, the RWTH Aachen University
built several Augmented Reality Microscopes. These microscopes are
capable of streaming their image to multiple devices and presenting annotations
and markers on their display. With the help of this technology, an
iPad application has already been developed. The application improves the
ability to give better feedback and instructions to the students by observing
the procedure on the iPad and creating real-time annotations for the microscope.
This system works in a local environment, where trainers and trainees
have to be in the same space.
This thesis provides an improvement on the system by integrating an
online e-learning concept into animal surgery. In the thesis a teaching and
learning platform was developed that allows students and teachers to connect
their Augmented Reality Microscopes, stream and annotate its image for
other users. The web application utilizes WebRTC for the transmission of
real-time media between the browsers, where the media travels through a
central media server. The media server enables the recording and managing
of conference calls and sessions. The application is simple to use and displays
the image of the microscope and user videos with low latency. Further, usercreated
annotations appear real-time for every participant.
Another important aspect of the thesis is the measurement of the overall
usability and user experience of the designed platform. For this reason, a
user study was conducted at the end of development phase to evaluate the
usability and measure the perceived Quality of Experience of the platform.
The result was an intuitive web platform based on the requirements de-
ned during the Requirements Elicitation. The platform enables teachers and
students to perform an audiovisual conversation, share their microscopes and
thereby teach and learn animal surgery online.

Monitoring and quantifying fatigue in team sports is a prominent procedure
among practitioners. Preventing athletes from experiencing excessive
levels of fatigue can prevent injuries, illnesses and increase success probability
of teams. Detection of fatigue in real-time is a challenging task that has
not yet been solved. The main goal of this work is to detect fatigue-induced
changes in player movement patterns during competitive games.
Using a Local Positioning System, players of an elite European ice hockey
team were tracked during all home games of their national championship.
Using the gathered data, rstly an in-depth analysis of player shifts was
performed. Changes in physical work performed during di erent stages of
shifts were evaluated. While it was hypothesized that relative workload during
the nal stages of a shift would decrease with an increase in shift duration
due to the accumulation of fatigue, the opposite was detected. Percentage of
total covered distance, total accumulated acceleration load and total accumulated
metabolic power produced during the nal stages of shifts increased
with an increased shift duration. Relative contributions during other periods
of shifts decreased with longer shift durations. Special player movements
(such as returning to the team bench) or fatigue have been hypothesized as
possible reasons for these ndings.
Further, positional data of players was used to compare a players movement
patterns with movement patterns of their teammates with the goal
of identifying fatigue-induced changes in player movements. In a rst approach,
player movements were compared using Dynamic Time Warping as
a distance-metric. In a second approach, player movements were compared
using a sliding window approach. Here, Mahalanobis distance was used to
quantify the abnormality of a single measurement during a window of 550ms.
Both approaches were able to identify anomalous player shifts, in which a
player displayed markedly di erent movement patterns compared to their
active teammates. However, when comparing metrics quantifying movement
intensities between anomalous and normal shifts, no signi cant di erences
were discovered. It was hypothesized that players of di erent positions (defensemen,
winger forwards and center forwards) move inherently di erently
during competitive games, rendering small-magnitude changes (as they may
be induced by fatigue) undetectable.
Further research regarding fatigue in ice hockey players is required. Analyzing
changes in movement patterns using internal load variables (such as
heart rate measures) appears to be an especially promising approach.

The electrophysiology of the heart is influenced by a variety of factors, including diseases, medication, and fitness. Particularly, cardiovascular dis- eases are detectable as anomalies in electrocardiograms (ECGs). Health problems can thus be detected by analyzing the recording of the electrical activity of the human heart. The resulting data can be fed into a machine learning (ML) model to detect anomalies. Traditional ML models require the data to be stored centrally. However, this poses the risk of violating data privacy, which is of utmost importance when it comes to sensitive data. One possible approach to address this challenge is federated learning (FL), which has gained significant attention in industry and research since 2016. FL of- fers privacy-by-design, as the private data of users remains on their devices and is not stored externally. Instead, the ML model is broadcast from a central server to the users’ devices, trained locally on private-held data, and the changes to the model are then sent back and aggregated on the server to form a new model. This process is repeated until the model meets predefined criteria, or can be a continuous process.
This thesis introduces an unsupervised federated ML approach using an autoencoder (AE). The objective is to obtain latent feature representations that facilitate detecting clusters of anomalous ECGs. A convolutional au- toencoder (CAE) is trained both in a federated and non-federated setting with data from two publicly available ECG datasets. The resulting latent fea- ture representations are visualized with principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE). k-nearest neigh- bors (kNN) classifiers are used to evaluate the ability of the feature encodings to differentiate between normal and anomalous ECGs. To a certain extent, encoded ECGs can represent beat classes well. However, types of rhythm are harder to distinguish. The performance for the two datasets is dissimilar, in particular for rhythm. A greater degree of discriminability between classes is achieved with the non-federated models than with their federated coun- terparts. The best performing models obtain a balanced accuracy of 84.06% with the federated and 85.26% with the non-federated approach for the beat classes of the MIT-BIH Arrhythmia Database.

Introduction: In the field of public health, behavioural aspects are commonly described
as determinants of health. These determinants have a major impact on one’s
health and well-being. With the rise and progress in the field of consumer-grade wearable
devices, individuals today have the opportunity to monitor part of their health
data, including behavioural determinants, themselves. The aim of this thesis is to
examine the impact, the use of wearable devices, such as smartwatches, has on
health-related behaviour and awareness in university students enrolled in German
universities. If the sole use of wearable devices causes people to improve their behaviour
and awareness in health-related matters, wearables will play an important
role in health promotion.
Methods: A quantitative survey was conducted, containing questions regarding behaviour
and awareness in health-related topics. The topics were: physical activity,
nutrition, sleep, stress, sedentary behaviour, substance abuse, and health-behaviour
on a broader scope, taking aspects such as utilization of offers by healthcare providers
into consideration. Additionally, wearable users were asked questions regarding
their usage of wearables. 101 participants completed the online survey. Users (n =
37) and non-users (n = 64) of wearable technologies were compared.
Results: Users and non-users did not report significant differences in behaviour and
awareness regarding physical activity, nutrition, sleep, stress, substance abuse and
health-behaviour on a broader scope. In awareness regarding the consequences of
a predominantly sedentary lifestyle, the group of wearable users reported significantly
higher scores but didn’t report significant differences in actual sedentary behaviour.
The user group reported a significantly higher discrepancy between awareness
about nutrition and their actual dietary behaviour than the non-user group. Both
found differences showed only a small effect size. When it comes to wearable-related
behaviour change, the users who successfully achieved behavioural changes or are
willing to change their behaviour are the dominant group only in the fields of physical
activity and sedentary behaviour.
Conclusion: The findings in this study do not support the proposition that the sole
use of wearables improves student’s health-related behaviour and awareness. Little
evidence, that indicates an impact on awareness on sedentary behaviour, and an
impact on behaviour regarding physical activity and sedentary behaviour was provided.
Even though students, who reported to be willing to change their behaviour
with the help of wearable devices were more likely to report higher scores in achieved
behavioural changes through wearables, the relation effect is not attributable to the
sole use of wearables. Further research, investigating if the sole use of wearables
improves students’ health-related behaviour and awareness, to draw conclusions for
health promotion, is needed.

Passive, non-intrusive medical condition detection systems have the po- tential to detect serious illnesses at an early state when being used in every- day situations. However, a large amount of privacy-sensitive data, which is not freely available, is required to develop such systems. This prevents the development of such systems.
We conceptually propose a way to mitigate this challenge by introduc- ing a general method of developing detection systems for certain medical conditions in a highly privacy-preserving manner. The method is based on the Secure Aggregation protocol, proposed by Bonawitz et al. in ”Practical Secure Aggregation for Privacy-Preserving Machine Learning”. It can aggre- gate individual, client-held values without revealing them directly to anyone involved. The lack of libraries allowing for the protocol to be used on iOS motivated this thesis to design and develop a new framework implementing the protocol.
As a first step, we motivate the need for a privacy-preserving vision- impairment detection system, which serves as an example for the proposed generic medical condition detection system. Secondly, the Secure Aggrega- tion protocol is explained in detail, as it constitutes the foundation of our proposed system. Thirdly, we describe our implementation of the protocol for iOS. Finally, the medical condition detection system is introduced and applied to the use case of detecting vision impairments.
To develop the system, we introduce a new way of using Secure Aggre- gation to calculate various statistical measurements about some client-held values in a privacy-preserving manner. This kind of information is used to establish expert systems capable of detecting medical anomalies. They are distributed and evaluated completely on-device, allowing for privacy during usage.

Die größte Mehrheit der Pflegebedürftigen in Deutschland wird durch die informelle häusliche Pflege versorgt. Pflegende Angehörige übernehmen die Aufgabe der Ver-sorgung dabei häufig mit dem Gefühl, weniger gut unterstützt zu sein und mit dem Wunsch, mit mehr Informationen zur Pflege ausgestattet zu sein. Im Zuge dessen wurde durch die Techniker Krankenkasse ein digitaler Pflege-Coach entwickelt, der informell Pflegende mit Pflegewissen ausstattet und den richtigen Umgang mit der neuen Rolle lehrt. Im Rahmen dieser Forschungsarbeit wird der TK-Pflege-Coach be-züglich seines Inhalts und seiner Konzeption analysiert, um Anregungen und Opti-mierungsvorschläge zur Weiterentwicklung des Online-Coachs abzuleiten.

In Form eines empirischen Forschungsansatzes wurden anonymisierte Umfragen, die mit den Nutzern des TK-Pflege-Coachs durchgeführt und der vorliegenden For-schungsarbeit bereitgestellt wurden, analysiert und hinsichtlich möglicher Optimie-rungsempfehlungen für den TK-Pflege-Coach interpretiert. Zunächst wurden dafür die Antworthäufigkeiten der Umfragen quantitativ mittels Python erhoben und aus-gewertet. Im Anschluss daran wurden alle erhobenen Textantworten der Umfragen durch eine inhaltlich strukturierende qualitative Inhaltsanalyse mit deduktiver Katego-riendefinition in der Software MAXQDA untersucht.

Die quantitativen Auswertungen der Umfragen zeigen, dass die Inhalte der Module des TK-Pflege-Coachs den Nutzern teilweise Anregungen zur Verbesserung der Umgebung des Pflegebedürftigen geben können. Zudem empfinden die Nutzer des Coachs die Fragen und Aufgaben überwiegend als nützlich. Die Ergebnisse verdeut-lichen überdies, dass einige Inhalte des Online-Coachs den Nutzern bereits bekannt sind bzw. dass die Nutzer in Bezug auf vereinzelte Themen des Coachs kein neues Wissen erlangen. Darüber hinaus ergibt die qualitative Auswertung der Textantwor-ten einige positive und negative Aspekte sowie Optimierungsanregungen zum TK-Pflege-Coach.

Aus der Interpretation der Ergebnisse gehen einige dringende sowie optionale Opti-mierungsempfehlungen und konkrete Umsetzungskonzepte hervor, die künftig zur Weiterentwicklung des TK-Pflege-Coachs beitragen können.

Die formulierten Optimierungsempfehlungen werden anschließend nach ihrer Dring-lichkeit geordnet und zusammenfassend dargestellt. Abschließend wird dabei aufge-zeigt, wie die Verbesserungsanregungen zum TK-Pflege-Coach vor einer Implemen-tierung auf Wirkung und Sinnhaftigkeit überprüft werden könnten.

Many trainees who practice strength training regularly log their workouts to record progress and plan future training sessions based on past ones. Since accurate recording of training-relevant parameters by hand is time-consuming it seems desirable to do this automatically during training with technical aids. Some work in the field of Human Activity Recognition (HAR) has already addressed endurance activity recognition, but there is a lack of research,
literature, and experience in single-sensor strength exercise tracking. In this work, we investigate the tracking capabilities of a single wrist-worn sensor, namely an Apple Watch for detecting common strength exercises. We pursue a machine learning approach and build a Long Short-Term Memory (LSTM) model using Apple's CreateML software and a 2D Convolutional
Neural Network (CNN) by leveraging the Keras framework on the Tensorflow backend. Our results show that the model created using Keras significantly outperforms the CreateML model due to the customization capabilities of the network parameters. For the twelve different strength exercises studied in this work, we achieve a classification accuracy of 99.16%. We further investigate the capabilities of automatic repetition counting during the execution of an
exercise. Using peak detection, we first analyze which sensor data available from the Apple Watch is crucial for the repetition counting of each exercise.
From our findings, we develop our own peak detection algorithm and achieve a combined mean absolute error of 1.09 repetitions per set and a mean relative error of 7.74%. In the course of this work, we further improve the overall
usability of the existing iOS fitness app SmartFit and extend the workout feature by integrating the created machine learning model for automatic exercise detection. We evaluate the usability using the System Usability
Scale and achieve a final score of 90.875 out of 100 possible points. Our results show that automatic exercise recognition during a workout is possible and most of the strength exercises studied in this paper are well recognized with the approach followed.

Dietary habits play a vital role in the overall health and wellbeing of an
individual. Consuming the wrong type of food can be as harmful as overeating
or undereating. Diffrent people can respond differently to different types
of food items, depending on various factors such as lifestyle, metabolism, age,
etc. The reaction of a body to certain food items can lead to various minor
or major health issues. Food Intolerance is a prevalent issue in today's society
and a affects a large percentage of the population. It is estimated that
the more common food intolerances such as lactose intolerance can affect
upto 65% of the total population. Moreover, while clinical tests are available
for the diagnosis, food intolerances are more commonly self-reported or
self-perceived.
In this thesis, we propose an algorithm based on statistical data analysis
to recognise patterns and correlations between food and health. The algorithm
identifes if any food item triggers health related issues in an individual
and also calculates the probability of having a food intolerance. We present
the solution as a mobile application called LactoCheck. The app tracks the
daily food consumption and the observed health issues for a person, and displays
the analysis based on the input data. LactoCheck mainly focuses on Lactose
Intolerance. The app informs the users if they might be suffering
from lactose intolerance and which food items might be causing them health
issues.
We carried out the project in three phases. The first phase involved
research on existing apps and current work in the field of nutrition and food
intolerances. The second phase consisted of planning and designing of the
application modules, followed by the development of the mobile application.
The third phase focused on testing and a user study. The algorithm was
tested in a simulated environment, while a user study was carried out to
evaluate the usability or user-friendliness of the mobile app.

Human action recognition has a wide range of applications, including
video surveillance and security, human-computer interaction, robotics, and
health care. Nowadays, skeleton-based action recognition has drawn increasing
attention thanks to the availability of motion capture devices and realtime
skeleton estimation algorithms in smartphones. We achieved human
action recognition on an iOS device using two approaches. Each of them
used a di↵erent framework from Apple to extract skeletal data. The first
approach extracted two-dimensional skeletal data from RGB videos using
Vision and then used the action classification option in CreateML to build
the model. The second approach extracted three-dimensional skeletal data
directly from the iPhone’s camera using ARKit and then used Keras to build
an LSTM model to classify the movements. We could not compare both
approaches directly, but our results suggest that the second approach using
ARKit and an LSTM model performs better than the first approach as it
classified more actions with smaller datasets.

Aim Since coordinative performance is closely related to sporting activity and the latter is
a major topic of prevention measures in Germany, further investigation of this relationship
and the variables influencing coordination is important. The main goal of this work is
therefore to identify possible factors influencing the coordinative performance of German
school children. In order to make the results comparable with other work and to make
them further usable in the future, Common Data Models are now frequently used for the
standardization of study data. Therefore the extent to which a transfer to the OMOP
Common Data Model is possible for a domain expert without IT background is discussed.
Methods The examined data originates from three test points of two evaluations of
the prevention measure Skipping Hearts, which is carried out by the German Heart
Foundation. Four coordination tests were conducted for each evaluation of coordination
performance. In order to prevent a distortion due to a learning effect, only the first test
participation was considered for each child. After data adjustment, the sample consisted
of n=2447 children. The observation and evaluation of the data was carried out with
the open-source software KNIME Analytics Platform. The transfer of the data into the
OMOP CDM was then planned by using the software Rabbit-In-A-Hat. Results The age
of the considered sample ranged from four to 16 years. The gender distribution was nearly
balanced. A positive correlation with coordination performance was identified for age,
height, weight, body fat percentage, BMI, hip and waist circumference and blood pressure.
Waist to hip ratio (WHR) and heart rate were negatively correlated with coordination
performance. The influence of gender varied by exercise. The data from the evaluation
process was imported into the Rabbit-In-A-Hat program for the CDM transfer. Here, the
data was assigned to four existing CDM tables. The calculated correlation data is not
yet supported in the OMOP CDM and must be stored as new concepts under the newly
created table CORRELATIONS. Conclusion The coordination performance and also its
correlation with anthropometric parameters depends strongly on the coordination exercise
under consideration. The positive correlations between coordination performance and
BMI, weight and body fat percentage require further differentiated investigation, as they
contradict previous studies. For the classification of a child’s performance, the next step
would be to establish reference values for the tests under consideration. Planning the data
transformation into the OMOP CDM was time-consuming and demanding. Therefore, it
has to be weighed up whether a research project is of such relevance in the future that
this effort is justified. Due to the necessary adjustments to the model, it should also be
investigated whether another CDM can better represent the data.

Gathering enough usable data in order to produce meaningful and sensibl results is, especially in a medical context, one of the biggest impediments
in machine learning. In this thesis, we utilise recurrent conditional GANs to generate labeled time-series data for data augmentation. For that reason,
we compare four architectures with different loss functions. We test the traditional GAN, Wasserstein GAN (WGAN), WGAN Gradient Penalty
(WGAN-GP) and propose a new architecture called Direct Lipschitz WGAN (DL-WGAN). We evaluate the models' performances on toy-datasets and on
real EMG data which was collected by a Thalmic Myo armband. By using a technique known as
"Train on Synthetic, Test on Real" (TSTR) we evaluate
DL-WGAN's generative quality on real EMG data. In the context of TSTR, we train a neural network classiffier with synthetic data from the DL-WGAN
and an identical classiffier on real data and compare their performance on new real data. The classiffier trained on fake data was able to reach an accuracy
of about 80 % even on datasets where the classiffier trained on real data could not learn the classiffication. This result indicates that the performance of machine learning models handling time-series data can be increased by using synthetic data from GANs.

In this paper a low-cost calibration setup for accelerometers and gyroscopes is considered. Smartphones, equipped with accelerometers and gyroscopes, are to be used as low-cost Inertial Measurement Unit (IMU)s for medical studies. In comparison to the shown setup there are industrial rate tables, which are not affordable for many studies. At first a record player was modified, so that the rotation speed can be controlled by the soundcard of a computer. The rotational speed is measured using an Infrared (IR) sensor and the reflectors on the side of the turntable platter. The measurement data from the IR sensor is recorded through the soundcard. Additionally a website was written, to access the IMU data of the smartphone in real-time.

Falls caused by balance problems are dangerous and can lead to severe consequences. To avoid such falls, it is important to recognize how strong and where the balance is compromised.
In order to find out whether balance problems can be detected by devices that are easy to handle, easily available and inexpensive, the Nintendo Wii balance board, the Noitom Perception Neuron motion capture suit and the Texas Instruments CC2650STK SensorTag are compared to each other with regard to their ability to provide information about one person’s balance.
The Wii balance board is a low-cost device that can be used to recognize how well the balance of a person works while standing. It can recognize how much the weight distribution of a person changes over a certain time while standing and has already been validated to give reliable center of pressure measurements. The motion capture suit provides information about the position of the body segments the sensors are mounted at. It has been compared to the Wii balance board to find out if it can also assess balance. Finally, the SensorTag is used as a cheaper alternative to the motion capture suit to measure the acceleration at the hip.

In physiotherapy, Proprioceptive Neuromuscular Facilitation (PNF) movements are
controlled movements performed by a physiotherapist on a patient to increase strength,
range of movement or
exibility. Learning to perform these movements accurately
requires repeated practice on the part of the physiotherapy student and appropriate
feedback on the part of the physiotherapy teacher. In currently established classroom
environments, the amount of feedback a teacher can provide is limited by time
constraints. An eLearning application that generates automated feedback when a
student performs a particular PNF movement can provide a solution to this problem.
A system to generate such automated feedback has already been proposed where
two Thalmic Myo armbands are placed on a student's forearms while he performs a
movement, and the recorded signals are evaluated against a trained Hidden Markov
Model (HMM) for that movement to determine the accuracy and detect errors. The
existing system gives an overall accuracy score for the movement execution and
displays the readings from the acceleration, gyroscope, orientation and surface Electromyography
(sEMG) with respect to the acceptable ranges for these values for a
correct execution.
In this thesis, this existing system is to be extended to an end-to-end application
that takes the sensor data from the Thalmic Myo armbands and displays easy to interpret
visualizations of the evaluation results on the student's smart phone. Along
with the overall accuracy score of the movement, the application provides feedback
by displaying points in time over the execution where errors occur and additionally
indicate the magnitude of the errors.

Electroencephalography (EEG) is a noninvasive and low-cost way to record brain activities. It is very useful to find information about brain-related diseases and disorders such as epilepsy, schizophrenia or sleep disorders. Analyzing EEG is timeconsuming and is usually based on the subjective judgments of clinicians.

In this thesis, following the work of Roy et al. [1], we present a system to classify whether an EEG is abnormal or not. The model takes the raw EEG data as the input and classifies it automatically. No expert knowledge about the domain is needed. The key concept of the system is the Recurrent Neural Network (RNN). The system firrst stacks multiple inception modules (formed by convolutional layers) followed by Gated Recurrent Unit (GRU) layers which are connected densely. The softmax layer
takes the output of GRU layer as the input and outputs the prediction result. Wereport an averaged accuracy of 85.94% for the abnormal EEG classification on the evaluation set of the TUH Abnormal EEG Corpus.

As the TUH EEG Corpus has a large amount of data that is unlabeled, we investigate whether it is possible to use the unsupervised pre-training method to improve the classifcation results. Based on our approaches and experiments, the unsupervised
pre-training method does not improve the performance of the abnormal EEG
classification task.

The distribution of predicted probabilities for each class is plotted in the thesis. From the distribution, we find that some samples in the dataset may be mislabeled.

Electroencephalography (EEG) is a non-invasive and relatively low-cost method for medical diagnostics and neuroscientific research. While it is useful to get information about different encephalopathies such as epilepsy, actually analysing an EEG is time-consuming and prone to human error. Tools for improving and accelerating the interpretation process are unambiguous visualizations and automated classification systems. For such a system an EEG can be modelled by the time series data as captured by the electrodes and the electrode placements on the scalp. Both informations are relevant when interpreting an EEG.
This thesis presents a method for automated classification of normal and abnormal EEGs. Because a convolutional neural network (CNN) is trained on labelled examples given by the TUH Abnormal EEG Corpus no specific knowledge about the domain is required. It uses the electrode placements and the time series electrode data by first creating a visualization, which shows the distribution of spectral content of the EEG over the scalp for a specific time range using an RGB image. Producing such an image for each time interval creates a video which now gets a label by first classifying each frame and then using the single frame classification scores in a final decision rule.
The automated classification of the whole EEG into normal or abnormal works with an accuracy of 82.99% while each frame of the generated video itself can already be classified with an accuracy of 74.28%. By visualizing the weights of the CNN more insight on the features that lead to an classification as abnormal in the generated images is given.

A study conducted by the Registered Nurse Forecasting (RN4CAST) showed that
in Germany on average a single caregiver in a hospital has to take care of up to
12.4 patients. During a night shift this number can rise up to 19.8 patients. Among
others this shows that caregivers are often overworked and stressed. To combat this
a smart shirt was developed that can detect and warn wearer if they having a bad
posture. The current smart shirt uses sensors on the shirt which are connected by
sewed in cables.
This bachelor thesis aims to develop a new prototype with a decentralized system
where the sensors can be freely attached, optimally on prescribed spots, to any shirt
and can be used to detect the posture of the wearer. To achieve this we use Bluetooth
Low Energy (BLE) sensors with their built-in gyroscopes. Since these BLE
sensors include different kind of sensors, including temperature,humidity,pressure
and optical sensors, we also investigated the usage of these sensor data and their
usefulness in detecting bad posture and stress.
The aim of this thesis is to develop an application for Android devices that evaluates
and analyzes the sensory data and notifies the wearer if he is having a bad posture
or if he is currently under stress. In addition to that we intend to do a small trial
with a number patient to see if the sensor can really detect bad posture and stress.

Functional magnetic resonance imaging (fMRI) and scalp electroencephalography
(EEG) are non-invasive brain imaging techniques. fMRI offers millimeter-accurate
spatial resolution and second-accurate temporal resolution. In contrast, EEG’s
spatial resolution is limited while its temporal resolution is in the millisecond range.
These complimentary characteristics make integration of both data types attractive.
In order to investigate the influence of two monoamine oxidase A gene (MAOA)
genotypes on the human brain, resting-state fMRI imaging data was collected while
EEG recording was done simultaneously. The dataset of 61 separate scanning
sessions of 25 subjects was labeled as acute tryptophan depletion (ATD), intake
of selective serotonin reuptake inhibitors (SSRIs), and placebo. A previous paper
from Eisner et al. (2017) had identified an effect on brain connectivity when carriers
of one genotype were subject to the ATD condition when compared to placebo, using
seed-based correlation analysis (SCA) on only the fMRI data.
This thesis, in two parts, analyzed the integration of the available resting-state EEG
and fMRI data. The first part looked at the linear correlation of fMRI to EEG using
the framework of general linear models (GLMs); while the second part performed
classification of condition and genotype using gradient tree boosting on various EEG
and fMRI features and convolutional neural networks (CNNs).
The correlation analysis resulted in the highest correlation being found in the frontal
lobe area and the 12.5-16 Hz (low-beta) EEG band with a 2-second onset delay from
EEG to fMRI. Therefore, the thesis concluded that significant information lies in
the EEG signal which can be used similarly to fMRI data.
The second part has established that the SSRIs condition can be clearly distinguished
from placebo using a gradient tree boosting algorithm. The best performance of a
mean area under the receiver operating characteristic curve (AUC) score of 0.668
in 100 times repeated, shuffled, stratified 6-fold cross-validation was achieved using
fMRI connectome features. No improvement could be realized by adding EEG raw,
and spectral features. The classification of ATD from placebo and genotype has
bordered random chance, and hence no significant discrimination could be achieved.
Additionally, the thesis concluded that using CNNs on raw time-series did not
produce any generalization.
In conclusion, the thesis could confirm the hypothesis that EEG and fMRI are
strongly correlated, and connectomes work best as predictors for resting-state classification.
Future work should, therefore, follow up on using different connectome features and
more fine-tuned classifier architectures; ideally, an acquisition of a larger dataset
could help with problems of overfitting and generalization of the used classification
algorithms.

The lung is an essential organ which absorbs oxygen from breathing air and releases carbon dioxide. Several reasons can lead to a damaged membrane structure and thus a damaged lung. Due to its relevance, patients with Acute Respiratory Distress Syndrom (ARDS) are often treated in intensive care units. The parameters for mechanical ventilation are determined based on a blood gas analysis. However, the quality of treatment is at the earliest detectable after the next blood gas analysis or even after some iterations. Furthermore, other diseases or circumstances - e.g., specific heart problems - can lead to anomalies in the blood gas values. This thesis tries to support the diagnosis and treatment of respiratory failure with machine learning methods. Therefore, different approaches are put to the test and evaluated. Among others, data on blood gases and ventilation parameters are extracted from the MIMIC III database to train a Hidden Markov Model and a Generative Adversarial Network.
Moreover, we investigate the diagnostic capability of the presented machine learning algorithms for the detection of cardiac diseases. In order to detect cardiac compli- cations, patients are monitored with the electrocardiogram (ECG). Since the inter- pretation of ECGs is not a trivial task, this thesis deals with the question, whether machine learning approaches can support clinical diagnostics.

Phase-contrast microscopy is a proven technique of improving visibility of semi-transparent unstained specimens that are very hard to see in bright-field microscopes. It was shown to be possible to also digitally process normal bright-field images to recreate the phase-contrast look (Digital phase-contrast). The digital technique has the advantage that it works with any bright-field microscope, conventional phase-contrast only works in special phase-contrast microscopes. Digital phase-contrast allows the use of a bright-field microscope by the firm Acquifer, which is better suited for integration in automatic processes than other phase-contrast microscopes.
A digital phase-contrast method that was implemented with such this bright-field microscope was able to create images that look very similar to phase-contrast images. Because of bright halo-effects around the cell walls, the cells are well discernible from the background. The digitally treated images also appear sharper than phase-contrast images, and the digital phase-contrast works in a larger area of the wells on a microtiter plate, while phase-contrast does not work at the edges of a well. A shading correction also removes uneven background lighting such that multiple images could be seamlessly merged to create are larger field of view. For these reasons digital phase-contrast can serve as a replacement for standard phase-contrast.

Primary snoring can have a huge impact on the quality of life of snoring persons but also on their bedpartners, although it is not considered an illness. Because of that, the present clinical trial examines the effectiveness of a speech and language therapeutic treatment focusing on the strengthening of the velum compared to a holistic speech and language therapy that excludes exercises focusing on the velum on the snoring of male subjects with primary velar induced snoring. Eight subjects are included in the study and randomly allocated to either therapy group or control group. They pass eight weeks of therapy, including one unit at 45 minutes per week under supervision and five further days of self-contained training sessions. The effectiveness of the therapy is measured by using the Epworth Sleepiness Scale, the Pittsburgh Sleep Quality Index and an acoustic analysis that examines the number of snoring events, the frequency and the amplitude. Due to the low number of subjects the results are analyzed descriptively. While the group results of the therapy group indicate improvement concerning the daytime sleepiness, the sleep-quality and the number of snoring events, the results are ambiguous regarding the single subjects in the group. The group results for the control group are stable concerning the daytime sleepiness and the sleep-quality, while there is an improvement concerning the number of snoring events. In regard to the single subjects, the results are ambiguous in the control group as well. Because of that, it is not possible to draw a distinct conclusion so that further research including more subjects is necessary in order to evaluate the effectiveness of the speech and language therapeutic methods on primary snoring.

Wearable devices have been evolving over the last years. The integration of sensors
on them encourages scientists to use wearable devices more frequently as they are
cheaper and the purchasing accessibility is easier than other specialized devices. In
this work, we will use a gesture control armband called Myo. Myo is equipped with
a gyroscope, magnetometer, accelerometer and a set of electromyographic (EMG)
sensors. Along with the Myo we will use a cell phone which receives the signals from
the Myo. The focus point of this master thesis is to research different features based
only on EMG signals coming from finger movements. EMG is a signal measurement
of the electrical activity of the muscles and is one of the most common sources
of information used to study muscle function and neurological disorders. One of
the main reasons behind that is that we want to correctly detect finger movements
using only the EMG signals that are associated only with arm muscles responsible
for motoring the fingers. Also, the existing literature is limited on finger gesture
recognition using non-intruding device such as Myo [1]. The goal of this master
thesis is to apply the earned knowledge to Myo controlled prosthetic arms worn by
amputees that have lost some or all of their fingers or patients that have deficiency of
moving their fingers due to a stroke, tendinitis, or peripheral neuropathy. The fact
that each finger has a different degree of freedom (DoF) gives us the chance to detect
various finger gestures. The most trite ones which we will examine are the extension
and the flexion of a finger or a group of fingers. In particular, the thumb can be
flexed and extended (2 DoF). The rest of the fingers can be flexed and extended in
one direction (1 DoF). The need of a solid classification between these two different
kind of movements is imperative. What’s more, there can be more combinations such
as holding three fingers (thumb excluded) as a cluster and moving forward backward
the remaining free finger. The increasing degree of complexity of the different finger
gestures plays a significant role in order to have a more complete detection cover
over the whole spectrum of finger gestures. After the collection of the recordings,
we will introduce a preprocessing stage which will help us in extracting features
from the raw data. In this work, we will use a range of methods such as root mean
square (RMS), moving average (smooth), Short-Time Fourier Transform (STFT)
and a variety of discrete wavelet transforms. The collection of all feature extraction
vectors will then be passed into classifiers and classification errors from each feature
will be compared with each other in order to determine the best performance. In
the implemented framework we will examine the approach of an Artificial Neural
Network (ANN) classifier.

The advent of induced Pluripotent Stem Cell (iPSC) technology has enabled stellar prospects for personalized medicine. To realize these prospects, high quality iPSCs are needed in large quantity. Therefore, the StemCellFactory (SCF) was developed, a joint project funded by North Rhine-Westphalia. Its aim is to automate the iPSC generation process, facilitating the cost-efficient generation of high quality iPSCs. Here, one key component that was custom developed for the SCF is the automated high-speed microscope. Its purpose is to visually assess iPSC cell culture status, increasing stability of the iPSC generation process as well as quality of generated iPSCs. To accomplish this task however, sophisticated image processing algorithms are required that currently do not exist.
The development of aforementioned algorithms was the aim of this thesis. To this end, a requirement analysis was performed in a first step together with iPSC experts from the Life & Brain GmbH for all stages of the automated iPSC generation process. Requirements revealed the need for fast and accurate multi-class segmentation of large images. Based on literature review, this was seen to be reliably provided by Deep Learning (DL). Thus, a solution concept employing DL based multi-class segmentation was developed in the next step, and the DL based multi-class segmentation implemented.
Here, first all classes to be segmented were defined, followed by the development of a semi-automatic image segmentation chain. Both were subsequently applied to generate a training dataset. Afterwards, a convolutionary encoder-decoder architecture based on the U-net was trained on the prior augmented training dataset using linear class frequency correction and step-based Stochastic Gradient Descent. The trained model was then val- idated using a manually segmented test dataset as well as visual estimations given by iPSC experts. Model validation revealed good segmentation accuracy with a weighted F1 score of 0.815 and equal to better performance compared to visual estimates from iPSC experts using Intraclass Correlation Coefficient (3.1). Thus, the model was considered fit for deployment and run-time analysis performed. Run-time analysis revealed that with the Graphical Processing Unit (GPU) currently integrated into the SCF, the given run-time requirement is unlikely to be met. However, it was also shown that by upgrading the GPU, run-time requirements can be met in the future.
Overall, these findings validated the DL based multi-class segmentation as the key compo- nent of the solution concept. Together with additional algorithms outlined in but beyond the scope of this thesis, the developed solution concept is expected to completely reproduce and capture all visual aspects assessed by iPSC experts during manual cell culture, unlocking the full potential of the developed high-speed microscope.

The education of motor skills is nowadays supported by wearable devices in so called serious games. This innovation is also present in the medical and dentistry education.
A first interactive game for the alginate mixing process was created by Hannig et al. in 2011 called Skills-O-Mat. This game is outdated and reengineered to a new technology and a new interactive device. Therefore, a website is created which provides the functionality of the gameplay. The input device is changed to a Myo armband. The connection between the wristband and the JavaScript website environment is achieved through the Myo Connect software provided by Thalmic. Also, the evaluation algorithm is reimplemented to fit the new input data. Therefore, a Multilayer-Perceptron is implemented and trained with the data of a professional
dentist. This neural network recognizes the different gestures done during the mixing steps. With this new version of the game, it is published to a greater audience and is more portable. The scoring scheme reflects the quality of the mixed alginate and a scoring
about 50% cannot be tricked. Skills-O-Mat is reimplemented as web based serious game sensor-assisted by the Myo
armband. The quality of the mixed alginate is measured with the help of a neural network which detects the different gestures during the mixing process. Additionally, a real-time feedback to the detected gestures is given. Whether the new evaluation
leads to a better training effect is not reviewed yet, but a training effect is detected in a small volunteer evaluation.

According to a National Health and Nutrition Examination Survey conducted in the
United States of America, about one tenth of the American population suffers from
clinically relevant sleep disorders. Snoring is a common sleep-related breathing dis-
order which affects approximately 40 to 50% of the American population. Not only
does snoring pose a potential health hazard, but also it is often a cause of disharmony
in couple relationships. Sleep disorders and snoring can be reliably diagnosed in a
polysomnographic sleep study, performed in a sleep laboratory. Recently, it has been
investigated how mobile devices such as smartphones and wearables can be used for
sleep analysis. SleepyLab is a sleep monitoring android application with a plug-in
interface, which can be extended with new sensor devices and analysis algorithms.
We propose two algorithms for the SleepyLab application. A snoring analysis al-
gorithm processes the SleepyLab audio signal to detect snoring. A sleep event de-
tection algorithm detects sleep events, that is patient movements and episodes of
arousal, from heart rate and accelerometer data. We tested both algorithms using
data recorded with the SleepyLab application. For training and evaluation of the
snoring analysis algorithm, we used data from a speech therapy study that has been
recorded with the SleepyLab. A sensitivity of 0.787 and a specificity of 0.937 for
snoring have been achieved. We found that the algorithm is suitable for usage in
the mobile sleep laboratory. The results of the sleep event detection algorithm have
been manually assessed and interpreted. The manual assessment suggests that the
detection of sleep events is possible.

The Myo by Thalmic Inc. is an electronic sensor device worn on the forearm, which is able to
detect arm movements and hand gestures. The detection is based on evaluating sensor
data from an integrated inertial measurement unit (IMU) and eight surface electromyography
(sEMG) sensors located along the perimeter of the armband.
The detection quality of the arm movements and hand gestures is highly dependent on the
quality of the data delivered by the device’s sensors and the calibration method implemented
by Thalmic via their software. This existing calibration is solely based on the gestures
developed by the company itself, while its specifics and implementation remain unknown
and are therefore not alterable. It improves the detection quality through the execution of an
initial gesture and, optionally, by learning from the user’s multiple executions of the existing
gestures.
For these reasons the calibration is not suited for the development and improvement of
gestures, which are not developed by Thalmic.
This thesis proposes a method that aims to improve the detection quality not only of the
existing gestures but newly developed gestures as well, by preprocessing the data acquired
via a direct Bluetooth connection to the device, effectively bypassing most of Thalmic’s
calibration process.
The primary goal of this thesis is the calibration of user-parameters, such as device position,
rotation, and immutable influences such as skin moisture and body fat for the sEMG or axis
misalignment and scaling errors of the IMU.
The acquired data is tested for the existence, magnitude and effects of these errors, to
check for already implemented software and hardware calibration methods, that could have
been applied before the data is sent and thereby determine the errors and appropriate
methods for further calibration.
Additionally, an algorithm requiring the user's interaction is developed, attempting to resolve
errors due to changeable influences and thus further increase the quality of the sensor data
as well as the gesture detection directly. However, the algorithm should require as little user
interaction as possible, since user variance is a considerable source for errors.
The algorithm requires the user to provide an approximate forward vector, which is used in
conjunction with the gravity vector provided by the pre-calibrated accelerometer, to rotate the
IMU coordinate system into a user coordinate system.
Furthermore, the unique layout of the Myo, with its eight sEMG sensors evenly distributed
around the user's arm, requires a rotation of the armband around the arm as part of the
calibration procedure to achieve the best possible quality of raw EMG data. To determine
whether the rotation-correction can be automatically applied by a mathematical model, the
results of such a rotation are compared with a rotation applied by the user.
The final calibration method is evaluated on a set of self-captured gestures. Additionally, a
publicly available dataset is used, containing a total of 154 captures composed of fourteen
official basketball referee gestures, including five number indicating gestures, which are
performed by eleven subjects. The method is evaluated with a slightly modified version of a
solution (Laukamp, 2015) for gesture recognition in hand hygiene, which compares three
different approaches to gesture recognition: k-Nearest-Neighbors, Support Vector Machines
and an approach which combines Artificial Neural Networks with Hidden Markov Models.

Proprioceptive Neuromuscular Facilitation (PNF) is a special concept of treatment
in the field of physiotherapy. The stimulation of sensory receptors and muscles are
combined with supporting a specific patient’s movement execution by the therapist.
It is often used to treat neurological, traumatic as well as orthopedic disorders in
rehabilitation processes.
Due to the high complexity of PNF movements, teaching the therapy method to
students is time consuming and requires adequate and proper training. However,
only a limited amount of supervised training time is usually available in current edu-
cational settings. Still, individual feedback on the movement execution is important
for trainees to improve their PNF skills. Additionally, learning or improving move-
ment execution from descriptions in books is difficult. In recent years, interactive
eLearning approaches have become more popular because they motivate trainees and
can be integrated into existing scholar environments. Existing systems for movement
evaluation and analysis are large, complex and expensive setups using multiple cam-
eras or body sensors and are, therefore, not suitable for teaching environments. As
an alternative, smart wearable devices are inexpensive, do not hinder the student
and can be connected to today’s computationally powerful smart phones. Yet, these
wearables are limited in their sensory capabilities and the sensor data is error-prone.
Nevertheless, they allow for the construction of a lightweight, simple and inexpensive
system that is required for an eLearning application.
This thesis proposes a system using two wearable Thalmic Myo armbands for rec-
ognizing errors in the movement execution of a given PNF movement and providing
constructive feedback to the executing student. The thesis deals with three main
tasks: Firstly, it aims at finding well performing feature sets based on the recorded
data containing important information about the movement execution. Secondly, a
set of recorded“gold standard”PNF movement executions will be used to train Hid-
den Markov Models especially designed for modeling movements and dealing with
variations in both sensory measurements as well as variations arising from different
executing therapists and patients. Thirdly, the trained model will be used to provide
feedback in form of classification, scoring and identification of error sources. The
evaluation of the developed methods shows promising results in modeling capabilities
of PNF and basketball referee movements. The scoring allows to easily distinguish
between correct and erroneous movement executions. Furthermore, the identifica-
tion of error sources allows to detect time durations of errors and the features that
show abnormalities.
The developed methods show promising results for automated feedback generation
and still have a large potential of further development. Therefore, this work is a
major step towards the development of an eLearning application in physiotherapy.

Electroencephalography (EEG) is one method to investigate different mental health
problems, like dementia, depression, schizophrenia or post-traumatic stress disorder.
Based on studies and the rapidly expanding population, the numbers of people with
these diseases are expected to increase even further. Clinical EEG devices are stationary,
large in size, complex and expensive. Over the past years many commercial
mobile EEG devices have been developed. To perform a complete mobile EEG measurement,
corresponding apps on mobile devices, such as tablets or smartphones,
need to be developed. An app for conduction of experiments was already developed
in previous work.
In this work BrainLabP&V, which is a mobile EEG visualization and processing
framework, is presented. By loading a previous recorded EEG measurement, the
proposed system can visualize and process the data. Here, a visual programming
approach is used to ease the operability, while increasing usability and allows the
system to adapt to changing user requirements. This makes BrainLabP&V another
important step towards completely mobile brain research.

Collaborative scientific writing is gaining increasing importance as the number of published scientific paper rises as well as the number researchers collaborating on research projects. Tool support for collaboration on scientific projects is still in its infancy, and students, researchers and
professors are left with unsatisfactory software support for collaboration on research projects that leaves room for improvement. The lack of adequate software support motivates the question for possible ways of improving the current state of the art.
In this thesis, a user study researches the habits and needs of researchers performing collaborative scientific writing at two universities. From the results of the study, a set of requirements
is gathered for a Social Collaborative Research Platform. From the deducted requirements, a
design approach and subsequent implementation of the novel tool
starTeX
is conducted, aiming to improve the current state of the art of tool support for collaborative writing. The design
approach is based on research findings of Computer Supported Cooperative Work and takes
the research theory of collaborative writing and three core aspects of systems supporting collaborative writing into account: technical, organizational and social. Especially the social component of the software aims to further increase the capability spectrum of current software and
acknowledges the social aspect of collaborative writing process. Specifically, an approach of
integrating the novel tool with a recent business-oriented social network software is presented.
The presented software undergoes evaluation regarding the aspects of collaborative writing
by a group of students and researchers. The results of the user survey have shown promising
results and motivate further development of the proposed platform.

Medical brain examination is crucial for detection and evaluation of disorders like epilepsy. However, complex hardware and technical set-ups are often required in the traditional way. Mobile and effortless healthcare can provide improvements regarding accessibility and costs especially for places with weak infrastructure. Fur- thermore, it creates space for novel scientific research.
In this work, BrainLab, which is an mobile EEG presentation framework for brain research, is presented. Using a commercial mobile EEG, the proposed system al- lows to conduct ERP experiments by just using a mobile device like a tablet or smartphone and without sophisticated equipment or infrastructure.
Several approaches to examine the viability of our framework for conducting EEG experiments have been exploited and shown promising results. Thus, BrainLab is a fine step toward fully mobile brain research.

Medical brain examination is crucial for detection and evaluation of disorders like epilepsy. However, complex hardware and technical set-ups are often required in the traditional way. Mobile and effortless healthcare can provide improvements regarding accessibility and costs especially for places with weak infrastructure. Fur- thermore, it creates space for novel scientific research.
In this work, BrainLab, which is an mobile EEG presentation framework for brain research, is presented. Using a commercial mobile EEG, the proposed system al- lows to conduct ERP experiments by just using a mobile device like a tablet or smartphone and without sophisticated equipment or infrastructure.
Several approaches to examine the viability of our framework for conducting EEG experiments have been exploited and shown promising results. Thus, BrainLab is a fine step toward fully mobile brain research.

The demographic distribution has dramatically changed over the last decades, re- sulting in an older population. This creates a situation where more caregivers – including informal caregivers – are needed. Informal caregivers often face higher stress due to their personal relation to the caretaker and the fact that they are not trained in clinical tasks. This might lead to health problems of the caregivers them- selves, resulting from false posture or physical fatigue. To overcome these problems, a smart shirt to detect false posture and physical fatigue has been developed. This smart shirt requires an infrastructure to record and to analyze the data.
The aim of this thesis is to develop this infrastructure to record data from the smart shirt and analyze it regarding posture and stress. This infrastructure is realized as an Android application, which connects to the smart shirt, records, stores, displays, and analyzes the data. The analysis methods are focused on posture and stress to face the main health challenges of informal caregivers. The application combines the recording and analysis of the data regarding posture and stress. With this combination of recording and analysis the application provides a low cost and easily available solution for informal caregivers.
This thesis contributes design, development, and evaluation of the corresponding application. For the latter, we developed and conducted a user study in which the participants performed various tasks. This study was used to identify whether the data from the smart shirt is significant to posture and stress. Ultimately, the application enables informal caregivers to monitor their health regarding posture and stress in an easily accessible and low cost fashion. This contributes to protecting the health of informal caregivers during their tasks caring for someone else.

In Germany, 25 percent of the adults suffer from sleep disorders, which can have
dangerous impacts like causing accidents or are followed by further disorders. Many
examinations that help to detect sleep disorders can only be done in sleep laboratories,
or need expansive hardware for usage at home. A possibility to perform a
cheaper and long-time monitoring of patients are smart wearable devices (wearables)
like smartwatches, which are equipped with sensors for a minimal monitoring.
This thesis aims at the development of a mobile sleep laboratory. Intended is a
platform based on Android, which collects and combines data from different sensors
of different devices. These sensors shall substitute the devices that are used in a
sleep laboratory as good as possible. To design the sleep laboratory as modular
as possible, the different components for monitoring and presenting data shall be
exchangeable. This allows the platform to be used for later research.

Recent progress in computer vision produced methods that allow to measure human heart rate only using a usual webcam or smartphone camera. This procedures introduce potential advancements in the field of medicine. Current methods for quick heart rate estimations are either overcomplicated for a quick estimation or hygienically questionable. This thesis explores the possibilities of using a computer vision approach in order to achieve a quick and contactless estimation of a patients heart rate using Google Glass. The heart rate is computed by measuring small changes in the skin color, which occur with every heart beat due to differences in the blood volume in the vessels. The method shows that the analysis of the green components is sufficient to get a reasonable estimation. The implementation of the method has shown deficits in hardware and software if the method is to be run in real time on the Google Glass. Since the Glass needs a continuous face detection because of strong movement of the region of interest, the creation of faster methods for face detection or a strong improvement in computational power is necessary if this method is to be run in real time. Possible solutions for creating such applications that run not in real time are finally presented.

The media frequently report about hospital-acquired infections and the devastating effects.
Since the hospital staff are in direct contact with the patients, the hands of the caregivers
are the ideal germ carriers. The transmission of germs from patient to patient can be prevented
by a proper hand washing or hand disinfection procedure. However, the workers do not often
follow the hygiene guidelines due to a lack of time or convenience.
The preliminary work of Laukamp et al. carried out a study and tried to figure out the effectiveness
of hand hygiene training of prospective nurses. The result was that the trainees had
problems in the practice of the described gestures. Thus, these deficits could be an additional
risk of germ transmission. This work is a proposal to solve this problem of lacking practical
experience. The task is to develop a mobile application which offers the trainees an opportunity
to practice the individual gestures. Based on the work of Laukamp et al., we use concepts of
gamification to keep the motivation of the users on a high level constantly. So, our application
motivates the user to perform hand washing gestures and is able to assess the performed gestures.
Additionally, the user is rewarded for a good performance.
This work tries to answer two research questions: (1) Is the concept of gamification a good
approach to motivate the user, and (2) which implemented gamification approach is the most
promising in the described setup. To answer the questions, a user study was carried out to
evaluate the usability and the aspects of gamification. The findings of this study are used to
formulate tasks for the future work.

With the recent changes in the German Civil Code (Bu ̈rgerliches Gesetzbuch (BGB)) patients’ will information is prioritized above all. Health care professionals must al- ways follow it when taking decisions regarding patients’ treatment. Advance Care Planning (ACP) is a prominent approach for patients to state their preferences in terms of accepted (or denied) future treatments. The will information is prepared together with health care professionals and expressed in a written form, e.g., organ donation cards, living wills or health care proxies. Patients may also state their representatives, who may make health-related decisions on patients’ behalf, in cases when the patients are not able to communicate or are legally incompetent to take decisions. Till now clinicians performed a time consuming and error-prone will infor- mation acquisition procedure, which typically involves patients’ relatives or persons empowered by the patients. The procedure is not applicable in emergency or life threatening situations, when the timely acquisition of patients’ will information is crucial for the decision making process. Moreover, the provided information is often open to interpretations and not objective.
A new approach to the problem of secure storage and delivery of patients’ will in- formation can be offered by Electronic Health Record (EHR) systems. They create, process and store digital health data, such as patients’ will information or history of patients’ diagnoses. This new approach allows timely delivery of patients’ will information on demand in a secure manner. We examined such EHR systems de- ployed in Europe. As far as we know none of the systems offers mobile applications for access to or management of patients’ will information.
This master thesis examines an alternative EHR system, which solves the problem of secure storage and delivery of patients’ will information by means of innovative mobile applications and a storage and processing module. This system delivers valid and up-to-date patients’ will information in emergency or life threatening situations. It guarantees the privacy of patients and health care professionals by pseudonymizing their personal information. Moreover, the confidentiality and integrity of the will information is assured during its entire life cycle through deployment of state-of-the- art cryptographical algorithms and protocols. Patients’ health data is stored in a secure data vault. The access to and the management of the health information is realised through two mobile Android applications. Patients’ will information may also be validated. The deployed security controls ensure patients’ data sovereignty. Furthermore, the EHR system is compliant with the current German data protection and information security legislation.
Considering the time-critical nature of the application scenarios the performance of system’s information delivery and validation functions are tested in terms of execu- tion time, processor load, memory and battery usage. The capability of the system to continue providing will information by peak loads is also examined. Moreover, the fulfillment of system’s design specifications is evaluated. The test results confirm system’s legal compliance and demonstrate the significant performance potential of the system under typical or peak loads. Means to improve system’s user interface and to extend the provided functionality are also discussed.

In recent years gamification has become a part in many areas of our daily routine. In regard to our personal life, companies like Amazon or Runtastic can base their gamification approach on publicly sharing personal achievements and statistics to improve user commitment. In contrast, gamification concerning our work life has to satisfy much higher privacy demands. Since comparison is a key component for gamification, privacy protecting computations of system wide statistical values (for example minimum and maximum) are needed. The solution comes in the form of secure multi-party computation (SMPC), a subfield of cryptography. Existing frameworks for SMPC utilize the Internet Protocol, though access to the Internet or even a local area network (LAN) cannot be provided in all environments. Facilities with sensible measuring systems, e.g. medical devices in hospitals, often avoid Wi-Fi to reduce the risk of electromagnetic interference. To be able to utilize SMPC in environments with Wi-Fi restrictions, this thesis studies the characteristics of mobile ad hoc networks (MANET) and proposes the design of a SMPC framework for MANET, especially based on Bluetooth technology, and the implementation as a C library.
Since MANETs have a high probability for network partition, a centralized architecture for the computation and data preservation is unfavorable. Therefor a blockchain based distributed database is implemented in the framework. Typical problems of distributed systems are addressed with the implementation of algorithms for clock synchronization and coordinator election as well as protocols for the detection of computation partners and data distribution. Since the framework aims to provide distributed computations of comparable values, protocols for secure addition and secure comparison are implemented, enabling the computation of minimum, maximum and average.
Devices of diverse computational power will be used to verify the applicability for wearables and Internet of Things (IoT) grade devices. Also field-tests with a smart phone ad hoc network (SPAN)(20-50 nodes) will be conducted to evaluated real life use cases. In contrast, the security of the framework and attack scenarios will be discussed. In summary, this thesis proposes a framework for SMPC for decentralized, distributed systems.

It is common that people undergo a regular medical check-up. But very few of us examine their mental health, whereas it plays one of the main roles in our daily lives. This work will be a part of a bigger project with the purpose to develop a screening system for identifying mental disorders and neurological problems.
The main goal of this thesis is to develop a modular system which explores a combination of signal processing and machine learning techniques to analyse electroencephalography (EEG) data. The system is evaluated by a particular task of recognition of four emotions: fear, disgust, joy and excitement. Yet, the proposed platform is not bound to that specific problem and can be used in a wide range of applications.
There exist several results on emotion recognition by means of the same EEG device. However current project uses a mobile tool to perform the experiments and a different kind of stimuli to evoke the emotions.
An Emotiv Epoc headset with 16 electrodes is used to measure the EEG data from the test subjects who will be presented with the visual stimuli from the IASP database. Two different experimenting platforms for stimuli presentation and data acquisition will be used and evaluated: a stationary PC and a fully mobile tablet-based system recently developed in the institute.
The results of the current work show that the mobile tool is applicable for the aimed task. The initial recognition rate of 30% for four classes could be increased up to 44% by performed optimizations and parameter adjustments. The results are above the chance and promising, but the design of the experiments should be reconsidered and a greater dataset should be built.

Depression is the most prevalent clinical disorder and one of the main causes of dis- ability. This makes early detection of depressive symptoms critical in its prevention and management. This thesis presents and discusses the development of Psycholo- gist in a Pocket (PiaP), a mobile Health application for Android which screens and monitors for these symptoms, and - given the explicit permission of the user - alerts a trusted contact such as the mental health professional or a close friend, if it detects symptoms.
All text input electronically—such as short message services, emails, social network posts—is analysed based on keywords related to depression based on DSM-5 and ICD criteria as well as Beck’s Cognitive Theory of Depression and the Self-Focus Model. Data evaluation and collection happen in the background, on-device, without requiring internet connection.

In our time, mobile electronic devices represent one of the areas with the most rapid develop- ment. Smart phones are ubiquitous and offer a level of computational power to the consumer that has surpassed those of super computers of past decades. Complementary, there is a growing market of wearable devices that interact with the user’s smart phone providing various func- tionality and data obtained from integrated sensors. The availability of these cheap and novel devices motivates the question for professional use cases. This work aims to assess the potential of mobile wearable devices in a medical eLearning scenario. We lay out the foundations for an approach that uses the Thalmic Myo to improve hand hygiene training for medical profes- sionals. This could decrease training costs as eLearning applications offer the possibility for unsupervised training and could improve long-term hand hygiene proficiency. To explore the technical feasibility, a machine learning framework is implemented to judge the execution of manual gestures, we focus on those belonging to the internationally applied WHO hand hygiene procedures. In the implemented framework, we compare various machine learning algorithms and data processing routines to maximize gesture recognition accuracy. Additionally, we present a study in which nursing students perform a hand disinfection wearing the Myo armbands to analyze the acceptance of and gather design requirements for a wearable-based eLearning ap- proach to hand hygiene training. This experiment also serves as an evaluation of the current state of hand hygiene quality, realized by measuring the hand cleaning success using fluorescent photographs. Taking into account both cleaning success and gesture recognition reliability, we present an outlook on the further development and other possible areas of application.

Non-invasive visualization of structures underneath the skin and organ surface in open surgeries and interventional procedures may help in surgery planning and navigation. In surgeries, visualization of concealed critical structures before dissection prevents surgical complications and results in better clinical outcome. During my thesis, I will be working on a part of running project at Philips research to develop an innovative hybrid operating room. My contribution is to determine the applicability of multispectral imaging in enhancing visualization of underlying structure non-invasively based on their optical properties. These structures will be utilized in potential clinical applications at present patient’s motion compensation during surgeries is considered as one of these applications. In our approach multispectral camera with filters of specific wavelength in infrared range will be used and the detected spectral signature of subcutaneous structures. Experiments will be carried out on volunteer subjects to build spectral images database for regions of interest (Arm, forearm, hand, lowerback and upperback) Then collected spectral images will be processed toward extracting the most relevant information by developing new image processing algorithms. With appropriate visualization technique and image processing algorithms, the information may be used for planning navigation in critical care interventional surgeries.

There is a growing need for caregivers in Germany, professional and non-professional, as the population is gradually growing older due to healthcare improvements. Informal caregivers face many challenges, from insecurity due to improper training for the performed tasks and lack of funding, to overestimating themselves physically and psychologically. The mHealth division at the Department of Medical Informatics, Uniklinik RWTH Aachen is currently conducting research to unobtrusively monitor physical characteristics of caregivers in a clinical as well as a private setting. The goal is to warn the caregivers if they are a performing a task incorrectly and/or if they are over exerting themselves. This thesis will support the on-going research by creating a low-cost wearable device for caregiver support. A prototype electronic system to measure various physical indications of fatigue and tiredness in caregivers will be designed and implemented using hardware readily available on the market. Need for a new system arises as already established devices such as the ‘SenseWear Armband’ are extremely costly, because they have very specialized sensors to measure physical activity. Cheaper versions of wearable devices that measure physical activity focus mainly on exercise and activities such as running/jogging to measure energy expended by the wearer. The system proposed in this project will be a wearable device, powered by an ‘Arduino Lilypad’ board. The device will be housed in a sports vest, designed to be worn under everyday clothing. The device will record signals from various accelerometers and an audio sensor, which will be used to measure the breathing rate, heart rate and movement of the caregiver. These signals will then be collected and communicated wirelessly using Bluetooth to an Android smartphone. This project will not only help informal caregivers in developed countries and further advance the field of telemedicine, but might also have applications in third world countries due to its low cost assembly.

A novel approach for guiding visually impaired people with the help of smartphones and image processing techniques is being developed. The aim of the project is to build a robust and affordable tool to provide navigation in outdoor environments. This work is part of this endeavor, where the task of improving 3D models in the image-guided positioning system was explored.
In the current state of the system, the result of performing 3D reconstruction from images consists of several models. However, a joint model to provide navigation routes is required. High computational costs cause the necessity of storing and compressing 3D data without loss of precision. Our initial assumption is that the size of the point clouds can be reduced significantly by removing outliers. Two problems, 3D model joining and outlier removal were addressed.
We implemented an application module, which takes as input models containing sparse 3D point clouds and their corresponding camera locations, together with camera GPS data. The module reduces the number of outliers in the initial models and estimates a precise geographical position for each.
Separate models are aligned in the same coordinate space with the help of freely avail- able geographical digital models. Based on successfully registered models, a combined 3D model is built in order to provide navigational information. The method implemented works well for models with small initial GPS errors. It is also able to register models with an initial GPS error of up to 40m, if they contain sufficient structural information for alignment.
For outlier removal, two existing approaches were implemented. Additionally, an own method was proposed. The three methods were evaluated in terms of their influence on localization accuracy. We observed that outlier removal does not improve the positioning accuracy of the system. We also observed that outlier removal does not degrade the quality of localization either, when the number of points removed is not higher than 25- 30% compared to the original model. Thus, we confirmed our initial hypothesis, stating that we can maintain positioning accuracy while reducing the number of points in a reconstructed model.

Eine große Herausforderung bei der Verarbeitung und der Analyse von Langzeit- Elektrokardiogrammen (EKG) ist, dass die sich ändernde Herzrate zu einer unterschiedli- chen Länge der einzelnen Abschnitte der Zyklen führt. Dies macht Vergleiche der einzelnen Zyklen und die Erkennung von statistischen Zusammenhängen schwierig, da viele Cluste- ringverfahren eine konstante Eingangsvektorgröße benötigen. In dieser Arbeit wird ein neues Hilfsmittel zur Analyse dieser Zusammenhänge vorgestellt: die nichtlineare Zeitnormierung für EKG-Daten. Diese passt die Zyklen des EKGs an die Länge eines Referenzzyklus an. Dabei müssen unterschiedliche Verschiebungen und Streckungen der einzelnen Anteile des Signals berücksichtigt werden. Hierdurch kann eine statistische Analyse direkt auf den Signalwerten der Zyklen ausgeführt werden. Bisher existieren keine Veröffentlichungen, außer der im Rahmen dieser Arbeit entstande- nen, die sich mit Zeitnormierungen von EKG-Zyklen beschäftigen, weshalb hier neue Wege beschritten werden müssen. Zur Normierung der EKG-Zyklen werden zwei Verfahren eingeführt: eines auf Basis eines modellbasierten Ansatzes, das Analytical Model Matching (AMM) und ein modellfreies Verfahren, das Dynamic Time Warping (DTW). Zusätzlich werden zur Bewertung der Verfahren Bewertungsmaße erstellt, so dass die Ver- fahren quantitativ verglichen werden können. Diese beschreiben die Ähnlichkeit des nor- mierten Zyklus zu dem Eingangszyklus, sowie zu dem Referenzzyklus. Es werden Experimente mit EKG-Zyklen von gesunden und erkrankten Patienten durch- geführt. Diese werden anhand der Bewertungsmaße ausgewertet. So wird festgestellt, dass beide Verfahren zur nichtlinearen Zeitnormierung geeignet sind, sie jedoch bei den verschie- denen Experimenten unterschiedlich gut abschneiden. Daher muss je nach Einsatzzweck abgewogen werden, welches Verfahren eingesetzt wird.

Anti-psychotics are important for the treatment of psychosis. They have a number of side-effects. Some anti-psychotics can cause changes in the brain. This changes can be detected in EEG records. In our work we investigate the relationships between anti-psychotics and slowings. Slowings are a common anomaly in EEG records characterized by a lower frequencies in the EEG record. To detect intervals containing slowings in the EEG records two methods were implemented. The first one applies a threshold and the second one machine learning. They do not achieve a sufficient accuracy so we developed an algorithm to classifies whole records that contain slowings or not. This method archived sufficient accuracy. Additional to the classifier the medication information have been extracted from the clinical reports. Final the classified records and the indicated medication is computed to look for a relationship.

Medicine students are known for their ability to learn by heart. Through medical school, they learn how the human body and the organs work. So, the question arises if there is a way to improve the quality of learning. It is known that game-based learning can support and increase motivation to study. The idea is to create a game which will help to study medical expertise. Since games give feedback and let the players know if they succeed or failed they will learn while playing. This thesis will reimplement the OrganChallenge within the .NET Framework using the Microsoft Kinect camera. Additional game modes will be introduced, e.g. a multi-player mode to play against each other and a mode to detect anomalies in medical images. It will be evaluated if the OrganChallenge increases the enthusiasm to study and if there will be a learning success with test players. As conclusion the goal will be to examine if the OrganChallenge can be used additionally for medical students to study.

The electroencephalogram (EEG), a signal measurement of the electrical activity of the brain, is one of the most common sources of information used to study brain function and neurological disorders. Mobile EEG systems ease long term monitoring and can provide improvements regarding accessibility and costs, especially for places with weak infrastructure. However, it requires an expert to analyze EEG recordings, in order to detect abnormal activities that occur in a wide variety of morphologies and can share similarities to waves that are part of normal EEG or to artifacts. Due to the large amount of data generated by EEG monitoring systems a human visual inspection is time-consuming and inefficient. Nevertheless, EEG is an important tool that is used to monitor side effects associated with medical drugs as for instance clozapine.
Clozapine is an atypical antipsychotic drug that has been shown to be effective in treatment-resistant schizophrenia and patients with intolerance to other neuroleptic drugs. One of the most prevalent side effects are seizures. Additionally various types of EEG abnormality are observable in many patients. Generalized slowing has been reported as the most frequent finding, followed by interictal epileptiform discharges such as spike and sharp wave activities that are typically observed in epileptic patients between seizures. While some studies on clozapine treatment mention epileptiform EEG abnormality as an indicator for seizures, others report that seizures are not necessarily predictable by previous EEG changes. Hence, long term EEG investigations are important in analyzing the effect of clozapine dose on EEG and the relationship between EEG alterations and seizures.
This thesis explores automated methods for the detection of abnormalities in order to support neurologists with the analysis of EEG recordings. We organize the various forms of EEG abnormalities in clozapine treated patients to investigate appropriate detection algorithms. Thus, we examine methods for the detection of slow activity as well as approaches for the detection of epileptiform discharges.
To our best knowledge, currently no dataset with manually marked individual spikes and sharp wave discharges is publicly available. Therefore in the thesis a subset of the Temple University EEG Corpus that provide recordings categorized into normal EEGs and abnormal EEGs will be used to examine the developed pipelines.

Nosokomiale Infektionen stellen stets ein relevantes Risiko für Patienten dar. Alleine in Deutschland kommt es jährlich zu circa 400.000 bis 600.000 nosokomialen Infektionen, wovon etwa 10.000 bis 15.000 tödlich verlaufen. Dies bedeutet, dass bei rund 3,5% der Patienten zur Zeit ihres Krankenhausaufenthalts eine nosokomiale Infektion vorliegt. Diese Zahl hat sich in den letzten Jahrzenten kaum verändert. Auch durch die Einführung der „Aktion saubere Hände“ der WHO im Jahr 2008 wird deutlich, dass es sich bei der Verbesserung der Händehygiene im medizinischen Sektor um eine weltweit relevante Aufgabe handelt.

Es existiert eine Vielzahl von mobilen Systemen, welche die Compliance in Bezug auf die Desinfektion der Hände misst. Mobile Sensoren erheben Daten über die Quantität der Händedesinfektion und weisen auf „Hygienemöglichkeiten“ hin. Diese Sensorennetzwerke nutzen in den meisten Fällen RFID oder Bluetooth-Sender, welche an das Krankenhauspersonal sowie an den wichtigen Stellen des Krankenhauses angebracht sind. Die Systeme dienen allerdings nicht als Trainingsmittel, sondern fungieren sie als Erinnerung zur Händedesinfektion. Auch ermöglichen sie keine qualitative Beurteilung der eigentlichen Durchführung.
In dieser Studie möchten wir mithilfe des Sensor-Armbands Myo by Thalmic und der mobilen Anwendungssoftware IdealPure, welche vom Institut für Medizinische Informatik der RWTH Aachen entwickelt wurde, die Händehygiene von Mitarbeitern im Gesundheitswesen trainieren. Unter Nutzung der Applikation werden den Studienteilnehmern Lerninhalte basierend auf den Empfehlungen der WHO präsentiert. Die Studienteilnehmer können dann die WHO-Händedesinfektionsroutine üben, während sie die Sensor-Armbänder an ihren Unterarmen tragen. Die Daten werden analysiert und mit dem Goldstandard verglichen. Die Studienteilnehmer erhalten eine allgemeine Leistungsbeurteilung und eine Bewertung bezüglich der Durchführung einzelner Desinfektionsgesten.
Im Rahmen der Studie möchten wie prüfen, ob sich durch die Nutzung der mobilen Anwendungssoftware die Qualität der Händehygiene der Studienteilnehmer sowie deren Motivation zur Durchführung dieser im klinischen Alltag verbessert. Sollte es möglich sein, eine signifikanten Verbesserung zu erreichen, könnte dies ein neuartiger Ansatz im medizinischen Hygienemanagement sein. Es würde das Personal befähigen, unabhängig und anonym unter geringem Zeitaufwand die Qualität ihrer eigenen Händedesinfektion zu verbessern.

Electroencephalograms (EEGs) are an important diagnostic tool
in modern medicine. Since it is not always possible to get a patient to an EEG machine, a mobile EEG device would grant medical professionals a new method to get an EEG from such patients, especially in rural or resource-austere areas. Additionally, the use of mobile EEGs can push the boundaries of brain research in collecting data in remote locations which lack access to EEG devices available in clinics or laboratories.
The main challenge which needs to be addressed is to assess whether mobile EEG devices are reliable and comparable in signal quality to the traditional EEG machines used for patients. The main task of this thesis is the collection of EEG data with mobile devices from patients and the comparison of the obtained results to a clinical EEG on a quantitative as well as qualitative level. Therefore, the procedure of the mobile EEG data collection should be as equal as possible to the traditional procedure. We assume to find evidence that mobile devices for EEGs are a good alternative to the traditional ones and supply sufficient quality for initial screening purposes, and in further development full diagnostics.

A system for Electronic Health Records is one of the hardest tasks of our time in the medical field. A strong indicator is, that there is only one country that implemented a EHR-system over the full life extension of citizens (Estonia). Germany is working on a system (eGK, developer: Gematik). There are strong challenges. Identifying these and suggesting a problem-solving approach is the aim of project „LauteKarte“. To identify the ideal communication way and abstraction level „360-delphi“, a stakeholder selected Delphi study method, is developed at the IMIB health division. An EHR-System over lifespan can among others correct especially the mismatch in End-Of-Life-decisions between patients wishes and medical standard procedures.