Machine Learning–Based Classification Framework for Human Health Care Monitoring

Abstract

Body-worn inertial sensors enable continuous activity recognition for remote health monitoring, but models often fail on unseen people, and static postures (sitting and standing) are hard to differentiate. This paper proposes MSA-HybridNet, a multi-scale attention hybrid network for supervised, subject-independent activity classification. Its deep branch combines three-scale parallel convolutions, squeeze-and-excitation attention, a bidirectional GRU and temporal attention pooling; its lightweight branch encodes a 180-dimensional statistical–spectral descriptor with gravity-tilt features, and branch-level auxiliary classifiers and explicit gravity input channels are added. The framework is tested on UCI-HAR and MotionSense under a strict subject-independent, multi-seed protocol, with McNemar tests against the main re-implemented baselines and a comparison with eight published studies. On MotionSense, it achieves 95.42 ± 2.51% accuracy (97.10% five-seed ensemble), outperforming all re-implemented baselines (McNemar p < 10⁻³⁷ for every tested competitor) and the 94.17% previously reported for a CNN-BiLSTM. On UCI-HAR, it achieves 94.56 ± 0.41% (94.98% ensemble), outperforming the re-implemented CNN, BiLSTM and CNN-LSTM (p < 10⁻⁷) and several published deep models, but not the classical 561-feature SVM (96.20%) or the strongest published hybrids. Ablations show that the handcrafted descriptor branch contributes most. With 0.29 M parameters and about 2.5 ms CPU latency per window, the model suits wearable health-monitoring deployment.