A Feature-Driven Hybrid Deep Learning for Human Activity Recognition in Holographic IoT Systems

Human Activity Recognition (HAR) is an essential component of intelligent healthcare and holographic IoT ecosystems, supporting continuous monitoring and real-time behavioral assessment. Despite recent progress in deep learning, challenges such as data imbalance, limited feature generalization, and insufficient exploitation of multidimensional contextual dependencies continue to affect model performance. To address these issues, this paper presents a feature-driven hybrid framework integrating a Multi-Head Convolutional Neural Network (MHCNN), a Convolutional Block Attention Module (CBAM), a Bidirectional Dilated Long Short-Term Memory Network (BiDLSTM), and a Multi-Head Attention (MHA) mechanism. The proposed holographic architecture combines spatial, channel-wise, and temporal representations to enable comprehensive contextual learning. Specifically, MHCNN extracts multi-scale spatial features, CBAM emphasizes informative spatial and channel responses, BiDLSTM captures long-range temporal dependencies, and MHA selectively focuses on relevant temporal patterns. Experimental evaluation achieves classification accuracies of 98.49%, 99.88%, and 97.23% on the WISDM, MHEALTH, and UCI-HAR datasets, respectively. Corresponding Matthews Correlation Coefficient (MCC) and Cohen’s Kappa Score (CKS) values of 96.98%, 99.76%, and 94.46% further demonstrate the robustness and reliability of the proposed framework. These results indicate its suitability for real-time, context-aware monitoring in holographic IoT-based healthcare environments.