Diabetes mellitus is a chronic metabolic disorder requiring meticulous blood glucose regulation to minimize both acute complications and long-term vascular damage. Traditional glucose monitoring approaches—such as finger-prick tests and continuous glucose monitoring (CGM)—primarily support reactive interventions, often falling short in enabling proactive management. This study proposes a deep learning-based predictive framework for blood glucose level estimation using historical CGM data. The model’s performance was evaluated using standard metrics including Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and R-squared (R2) score. Experimental results across multiple patients reveal that the model achieved RMSE values ranging from 19.37 to 28.57, and MAPE values between 7.76 and 13.31. The highest predictive accuracy was observed for Patient 570 (RMSE: 20.38, MAPE: 7.76), while the model struggled with higher variability in Patient 559. These findings demonstrate the model’s potential in delivering personalized, anticipatory glycemic control, thereby supporting more effective diabetes management strategies.

This paper introduces an integrated Hardware-In-the-Loop (HIL) testing framework, combining the UVA/Padova Type 1 Diabetes Simulator with AndroidAPS, an open-source artificial pancreas system. This integration forms a testing environment capable of evaluating insulin regulation algorithms under both virtual and real hardware conditions. The FDA-approved UVA / Padova Simulator models glucose-insulin dynamics and meal digestion. Paired with AndroidAPS, the system can actuate real-world insulin pumps to test insulin delivery control algorithms. The framework is tied together by various REST APIs and uses the Flask framework for efficient data exchange and system connectivity. The HIL approach provides a robust platform for functional and reliability testing of these algorithms. The developed APIs are open-source: https://github.com/OE-Diab/AP-HIL

Automated brain tumor classification is an intensively investigated problem, which recently attracted significant attention. Convolutional neural networks (CNN) and deep learning represent the standard for the foundation of any recent solution. This paper proposes two simplified VGG architectures and investigates their capabilities and limitations, in comparison with state-of-the-art CNN networks deployed via transfer learning. Various parameter settings are involved in the evaluation process, including different kernel sizes, dropout rules, loss functions, etc. Networks are trained and tested on a public brain tumor classification data set consisting of 3064 images and three tumor classes (meningioma, glioma and pituitary tumor). The thorough evaluation process revealed that the proposed CNN models can achieve competitive performances with regard to state-of-the-art methods in several scenarios. The best achieved accuracy benchmarks are 98.2% overall Dice similarity score and correct decision rate, and AUC values over 99.6% for each of the three tumor classes.

Reinforcement learning (RL) has shown promise in controlling blood glucose levels in a personalized way in type 1 diabetic patients. In this study, we investigate the impact of different activation functions and layer numbers on RL performance in blood glucose control. We train RL agents with various combinations of activation functions and layer numbers on a virtual patient model. The RL agents are evaluated based on their ability to maintain blood glucose levels within a target range while minimizing the frequency and magnitude of hypoglycemia and hyperglycemia events. Our results show that the choice of activation function and layer number significantly affects the RL performance. Specifically, the agents with ReLU activation functions and two or three hidden layers outperform the other agents, achieving a higher percentage of time in the target range and fewer hypoglycemia and hyperglycemia events. These findings provide valuable insights for the development of RL-based blood glucose control systems in type 1 diabetic patients.

We present a deep learning model for finding human-understandable connections between input features. Our approach uses a parameterized, differentiable activation function, based on the theoretical background of nilpotent fuzzy logic and multi-criteria decision-making (MCDM). The learnable parameter has a semantic meaning indicating the level of compensation between input features. The neural network determines the parameters using gradient descent to find human-understandable relationships between input features. We demonstrate the utility and effectiveness of the model by successfully applying it to classification problems from the UCI Machine Learning Repository.

One of the most challenging area of diabetes research is to provide such automated insulin delivery systems – so called artificial pancreas systems – that have robust and adaptive capabilities in a highly sophysticated way. I.e. they are able to provide robust insulin delivery actions at the beginning of the therapy to satisfy the requirements of the patients without knowing the users daily lifestyle and preferences however adaptive on the short-term to learn these patient specifics to increase the quality of the therapy. One possible solution is the closed-loop systems that have self-learning features. In the present study, we have examined a glucose regulatory problem using direct reinforcement learning based controller. The approach represents the fully automatic insulin administration as the timepoint and the carbohydrate content of the meals were unknown and randomized. We constructed a virtual environment of the patient with type 1 diabetes by applying a mathematical model. Proximal policy optimization learning model with continuous action space was used. Furthermore, we evaluated the effect of different training lengths on the test scenario.