Hybrid CNN–BiLSTM–GRU architecture for PAMAP2-based human activity recognition in patient monitoring

Human Activity Recognition (HAR) is essential in modern patient monitoring systems, facilitating real-time health assessments and timely interventions. This study introduces a hybrid deep learning framework that combines Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory (BiLSTM), and Gated Recurrent Units (GRU) to achieve high-precision activity classification using the PAMAP2 public activity sensor dataset, which includes 18 activities collected from nine subjects. In this architecture, the CNN module extracts complex spatial features, the BiLSTM component captures bidirectional temporal dependencies, and the GRU layer improves sequential processing efficiency. Through thorough preprocessing and hyperparameter optimization, the model achieves an impressive accuracy of 99.50%, surpassing previously reported benchmarks on the dataset. This outstanding performance underscores the model’s suitability for reliable patient monitoring and its applicability to related areas such as elderly fall detection, fitness tracking, smart home automation, rehabilitation monitoring, and workplace safety. Future enhancements may include integrating attention mechanisms to highlight critical patterns, employing transfer learning to enhance generalizability, utilizing multi-sensor fusion for richer contextual information, and optimizing the model for lightweight deployment on edge devices to ensure efficient real-time responsiveness.