Cinar, Beyza

Hypoglycemia is a serious condition associated with increased mortality in patients with type 1 diabetes, which is an incurable autoimmune disease. Hypoglycemia is defined by blood glucose levels below 70 mg/dL. The causes can include excessive insulin injections, skipping meals, or increased physical activity. It can occur suddenly and may be asymptomatic, impeding timely preventive measures. Thus, innovative technologies, such as machine learning, can help to predict the state before it occurs. Prediction models are mainly classified as short- and long-term prediction horizons (PHs) of up to 2 hours and up to 24 hours, respectively. Most research conducted in the field of diabetes forecasts blood glucose values. Still, the obtained accuracy may not be su!cient to prevent hypoglycemia due to the possible time lag of CGM devices. Moreover, most studies focus on one PH only. This thesis included short- and long-term PHs in the same classification model to consider multiple use cases and to enable better decision support. The predicted times are 5-15 min, 15-30 min, 30 min-1 h, 1-2 h, 2-4 h, 4-8 h, 8-12 h, 12-24 h before hypoglycemia. The input features are prior glucose measurements, the administered basal and bolus insulin dosages, and acceleration data. First, a correlation analysis between the input features and the classes is conducted. Thereafter, RNN and CNN are explored to classify the onset of hypoglycemia based on the proposed nine classes. Furthermore, training with six classes classifying up to 4 hours before the onset is compared. Finally, subject-specific models are tested. The population-based correlation analysis reveals a very weak association between basal insulin and glucose, and between basal insulin and acceleration data. An individual correlation analysis showed stronger relationships, but the scores varied significantly among the subjects. For the classification model with nine classes, the best results are obtained with a LSTM model. Subject-specific models improve the performance. However, only classes 0-2 could be well classified with recalls of 98%, 72%, and 50%, respectively. A population-based model with only six classes obtains better results with recalls of 99%, 73%, and 56% for classes 0, 1, and 2, respectively. In conclusion, the proposed system that includes short- and long-term PHs is not feasible with the data or models used. Whereas, a model classifying multiple short-term horizons up to 4 hours before hypoglycemia produces promising results with improved precision, and F1-measure and indicates that at least 60% of events can be predicted which is increased to approximately 70% in subject 563.

Die automatische Erkennung von Emotionen mit Hilfe biologischer Signale ist ein sehr vielversprechendes Forschungsgebiet in den Gesundheitswissenschaften. Vor allem die fortschreitende Entwicklung von Wearables und Cloud-Computing ermöglicht eine kontinuierliche Erfassung der Daten und die Erkennung der Emotionen, welche helfen frühzeitige Diagnosen von psychologischen Erkrankungen wie Depression festzustellen. Ebenfalls können Therapiemethoden entsprechend dem psychologischen Wohlbefinden individuell angepasst/gestaltet werden. Gängige Sensordaten hierbei sind der Blutvolumenpuls, die Herzrate, die Herzraten-Varibilität, die Hautleitfähigkeit und die Hauttemperatur. Nach einer Filterung und (statistischen) Merkmalsextraktionen der Signale werden öfteres maschinelle Lernverfahren zur Klassifizierung benutzt (Support Vector Machines, K-Nearest-Neighbor, Random-Forest...). Neuerdings gibt es auch Forschungen für Deep-Learning Methoden wie Convolutional-Neural-Networks. Für die Emotionsklassifizierung gibt es zwei konkrete Emotionenmodelle, das diskrete, in welchem vorbestimmte Emotionen analysiert und das dimensionale, in welchem Emotionen als Kombination (Vektoren) aus mehreren Komponenten (Dimensionen) dargestellt werden. Hierbei ist das zwei dimensionale Model am gängigsten, in welchem eine Achse die Intensität und die andere die Polung der Emotion darstellt. In dieser Arbeit wurde mittels bereits durchgeführter Studien, welche das zwei dimensionale Model benutzt haben, unter Betrachtung der verfügbaren Sensoren analysiert, ob Wearables eine gute Basis für die Ermittlung von Emotionen darbieten. Die Analyse zeigt, dass Wearables vielversprechend sind und genaue Ergebnisse liefern können, jedoch müssen Daten sehr gut für die Klassifizierungsmethode vorbereitet werden. Zudem ist eine große Datenmenge und homogen verteilte Gruppe an Probanden notwendig. Es wird festgestellt, dass die Genauigkeit stark von den Probanden abhängt und Emotionen sehr subjektiv bewertet werden. Des weiteren scheint das vorgestellte zwei-dimensional Model nicht ausreichend zu sein und es wird eine Erweiterung vorgeschlagen, um bessere Grenzen zwischen ähnlichen Emotionen zu ziehen. Letztlich kann durch den Vergleich verschiedener Arbeiten angenommen werden, dass es nicht das genau Richtige oder die Beste Klassifizierungsmethode/Algorithmus gibt und für jede Datenmenge die beste Methode “erkundet” werden sollte.

Blood analysis is a major procedure to diagnose and monitor diseases, but conventional methods lead to delayed interventions. This work aims to analyze current advances in non-invasive blood analysis. The identified methods include optical sensors, image processing and biosensors. Image processing can be used for self-screening but is not reliable for clinical monitoring. Optical and biochemical sensors are promising and can be extended to more components of interest. More research in this area is expected in the future combining different components.

Patients with type 1 diabetes (T1D) have a higher risk of experiencing hypoglycemia, which is a severe condition of decreased blood glucose levels under 70 mg/dL and can result in coma, or death in the worst case. Prediction algorithms could improve diabetes care by enabling preventive actions, but research is limited by available multivariate datasets. Thus, this work investigates the feasibility of transfer learning between two different datasets of people with T1D and type 2 pre-diabetes using a 1 Dimensional Convolutional Neural Network (1DCNN) model. Moreover, different thresholds for defining hypoglycemia are compared for the pre-diabetes group. The results show that transfer learning could be feasible if the model is trained on T1D with a threshold of 70 mg/dL, while a threshold of 80–85 mg/dL achieved the best training performance for the pre-diabetes data.