Hybrid Self-Attention-Based Human Activity Recognition Model for Consumer-Centric IoHT Devices

Human Activity Recognition (HAR) plays a pivotal role in consumer-centric Internet of Healthcare Things (IoHT) applications, including wearable health monitoring, Parkinson’s disease management, and home-based rehabilitation systems. Despite advancements in deep learning (DL)-based HAR, challenges such as biased datasets, suboptimal feature extraction, and complex activity interpretation persist. This paper presents a hybrid self-attention-based HAR model specifically designed for consumer-centric IoHT devices, aiming to improve accuracy, reliability, and user experience. The proposed model combines multi-head convolutional neural networks (CNN), long short-term memory (LSTM), and bidirectional long short-term memory (BiLSTM) networks, leveraging self-attention mechanisms to enhance feature extraction, address data biases, and effectively capture salient features. Its multi-head and multi-kernel architecture improves abstraction and models transient dependencies, enabling precise recognition of complex and similar activities. Evaluations on benchmark datasets—WISDM, MHEALTH, and UCI-HAR demonstrate accuracies of 99.04%, 99.79%, and 98.18%, respectively. The comparative analysis demonstrates its superiority over state-of-the-art models, demonstrating its potential as a robust and accurate HAR solution for consumer healthcare electronics and wearable IoHT devices.