Real-World Medication Adherence Monitoring Using Hybrid Neural Networks and Smartwatches

Medication non-adherence remains a critical healthcare challenge, contributing to poor clinical outcomes and increased healthcare costs. While multiple factors underlie non-adherence, the lack of effective and unobtrusive monitoring solutions limits the ability to provide timely interventions. To address this gap, we present a pill-taking gesture recognition framework that leverages smartwatch inertial data as a scalable and non-intrusive sensing modality. A hybrid deep learning architecture is introduced, combining convolutional layers, squeeze-and-excitation blocks, attention mechanisms, and long short-term memory units to jointly capture spatial and temporal dependencies inherent in multi-step pill-taking behaviors. The framework is evaluated using a free-living dataset under a leave-one-subject-out cross-validation setting, ensuring robust assessment in real-world scenarios. Results show strong recognition performance, achieving a precision of 0.92 with an optimal window length o f 9 seconds. Furthermore, a lightweight domain adaptation strategy enhances inter-user generalization, delivering up to 19% improvement in cross-user evaluations. Extensive experiments across different window sizes and demographic groups confirm the model’s robustness, scalability, and fairness. These findings demonstrate the potential of wearable-based gesture recognition for real-time, privacy-preserving monitoring of medication adherence, paving the path for the scalable deployment of such systems in both clinical and home environments.