Trabalho em anais de eventoOpen Access

Smartwatch-based gesture recognition for medication adherence via CNN-BiLSTM networks

RYLO, MARCOS NEGREIROS; ALCANTARA, JOAO ARTHUR; AYRES, FLORINDO A. C.; DE MEDEIROS, RENAN LANDAU P.; DE BESSA, IURY VALENTE; DE LUCENA, VICENTE FERREIRA

2026 • 2026 IEEE International Conference on Pervasive Computing and Communications Workshops and other Affiliated Events (PerCom Workshops)

Resumo

Medication adherence is a critical determinant of treatment success, yet objective and unobtrusive monitoring remains a challenge. This work presents a smartwatch-based gesture recognition framework for detecting medication intake actions via inertial sensor data. We evaluate a deep learning model that combines one-dimensional convolutional layers with a bidirectional LSTM architecture (CNN–BiLSTM) and investigate the impacts of the temporal window size, the number of convolutional filters, and the number of LSTM units on the classification performance. Data were collected from 20 volunteers performing five medication-related gestures via a Samsung Galaxy Watch 5 at 100 Hz. Composite sequences were constructed to emulate real-world continuous actions and segmented into windows of 50, 75, and 100 samples. The results show that larger windows improve accuracy, achieving a test accuracy of 0.932 and a validation accuracy of 0.951. Notably, shorter windows (0.5–0.75 seconds) also delivered competitive performance while substantially reducing memory usage, enabling real-time inference on resource-constrained devices. We also propose an IoT architecture that integrates wearable sensing, smartphone edge inference, and cloud-based adherence monitoring. These findings demonstrate the feasibility of short-window CNN–BiLSTM models for ubiquitous, real-time medication adherence monitoring in mobile health applications.

Autoria