Publication Date
Spring 2026
Degree Type
Thesis
Degree Name
Master of Science (MS)
Department
Applied Data Science
Advisor
Vishnu Pendyala; Mohammad Masum; Shih Yu Chang
Abstract
Seizure risk monitoring with wearable physiological data is difficult due to temporal uncertainty, patient heterogeneity, class imbalance, and severe privacy restrictions. Existing techniques frequently rely on static supervised learning or intrusive EEG-based monitoring, which limits their usefulness for long-term, tailored, and privacy-preserving deployment. This thesis presents seizure risk monitoring as a sequential decision-making issue with reinforcement learning, allowing for temporally consistent and prevention-oriented choice policies. A tailored framework is created using a Proximal Policy Optimization (PPO) agent and a Transformer-based temporal encoder, and it is tested with extensive ablation studies on wearable-inspired physiological signals under severely unbalanced settings. The findings show that reward-driven learning and temporal consistency give more relevant evaluation than point-wise accuracy in safety-critical monitoring activities. The framework is further expanded to include a federated reinforcement learning context, which allows for collaborative learning across patients without exchanging raw data. To address instability caused by non-IID data and weak-signal patients, we offer a trust-aware aggregation technique that dynamically reweights patient contributions based on risk-state detection performance and reward alignment. This technique increases robustness and encourages balanced conduct among diverse patients. Overall, this study proposes a proof-of-concept framework for tailored, privacy-preserving seizure risk monitoring with wearable data, laying the groundwork for future scalable and patient-centred healthcare monitoring systems.
Recommended Citation
Kapadia, Mayank, "Privacy-Preserving Personalized Seizure Risk Monitoring Using Reinforcement Learning and Federated Learning on Wearable Data" (2026). Master's Theses. 5780.
DOI: https://doi.org/10.31979/etd.c8tu-m7jh
https://scholarworks.sjsu.edu/etd_theses/5780
