The management of Type 1 Diabetes (T1D) requires continuous monitoring and precise control of blood glucose levels, which can be influenced by... Show moreThe management of Type 1 Diabetes (T1D) requires continuous monitoring and precise control of blood glucose levels, which can be influenced by various physiological factors such as physical activity (PA) and acute psychological stress (APS). This dissertation presents a novel multivariable real-time detection system designed to identify PA and APS, enhancing the efficacy of artificial pancreas (AP) systems. Using data from wearable devices, such as the Empatica E4 wristband, various physiological signals were captured, including blood volume pulse (BVP), accelerometer data (ACC), galvanic skin response (GSR), and skin temperature (ST). These signals were processed to extract features critical for classifying PA and APS.
A Long Short-Term Memory (LSTM) neural network model was employed to classify different types of PA and APS events. Additionally, a multitask learning framework was developed to simultaneously estimate energy expenditure (EE) alongside the classification tasks. The study incorporated explainable artificial intelligence techniques, such as SHAP (Shapley Additive Explanations), to interpret the model’s decisions and ensure that physiologically relevant features were used in the classifications.
A real-time system was implemented, integrating the detection of PA and APS events into an automated insulin delivery (AID) system. This system was validated through real-time testing with participants, demonstrating its ability to respond dynamically to physiological changes and provide timely insulin adjustments. The models achieved high classification accuracy, demonstrating that the integration of PA and APS detection into AP systems can lead to more precise insulin delivery, thereby improving glycemic control in individuals with T1D. Show less
