This dissertation presents novel machine learning solutions for modeling and automating processes in healthcare, with a primary focus on diabetes management and medical image processing. In the diabetes domain, the work introduces methods to detect patients physical activity from both synthetic and real-world data, integrating continuous glucose monitoring (CGM) and heart rate signals. By employing conservative machine learning algorithms alongside more advanced recurrent neural networks (RNNs), the dissertation demonstrates how accurate exercise detection can be achieved even under real-life conditions thereby enhancing personalized feedback for patients. Beyond activity detection, the research also addresses the automated recognition of food intake gestures using wearable sensor data. This solution is crucial for improving dietary records in diabetes therapy, as it accurately identifies when carbohydrate consumption occurs. The dissertation further proposes a reinforcement learning (RL) framework for insulin regulation, wherein policy and value networks learn optimal dosing strategies based solely on CGM data. Experiments show that this RL-based controller robustly maintains blood glucose within safe limits, offering a foundation for more adaptive and patient-specific “artificial pancreas” systems. In medical image processing, the dissertation I developed advanced architectures for segmenting and classifying magnetic resonance imaging (MRI) scans. Specifically, it introduces refined U-net designs ranging from 2D to spatiotemporal 3D to segment infant brain tissues into cerebrospinal fluid, gray matter, and white matter with high accuracy. Furthermore, a specialized “L-net” approach combines a U-net segmentation backbone with a convolutional neural network (CNN) classifier to distinguish glioma, meningioma, and pituitary tumors. Extensive experimentation reveals state-of-the-art performance in MRI tissue segmentation and tumor classification, underscoring the potential of deep learning to enhance diagnostic precision. Collectively, these contributions highlight how machine learning-driven methods can automate critical aspects of clinical management and medical imaging. By uniting algorithmic innovations with real-world application needs, the dissertation underscores the transformative potential of data-driven approaches in modern healthcare.