Multimodal healthcare system for human activity recognition using multiple features and advanced ensemble classifier

Human activity recognition (HAR) is applicable in various areas as it helps in healthcare monitoring, exercise assessment, and monitoring of smart devices.

To achieve high recognition, the study makes use of the synergistic integration of sophisticated signal processing and classification methods. First, we applied fourth-order median filtering and Hamming window processing to sensor signals, preserving activity-related changes while reducing excessive noise. Next, we extract several features, such as Shannon entropy, mel-frequency cepstral coefficients, spectral energy, spectral centroid, spectral flux, and dominant frequency, which enable us to gather information from both time and frequency domains. Subsequently, we adopt quadratic discriminant analysis to select the strongest features, facilitating easier identification of different classes. The final step involves training an ensemble of multi-layer perceptron (MLP), sparse MLP, and spatial-temporal MLP models, with all predictions made by each model combined through soft voting.

The proposed method demonstrates exceptional performance on three benchmark datasets, PAMAP2, Mobile Health, and Heterogeneity Human Activity Recognition, with accuracy values exceeding 95%.

The results clearly illustrate the effectiveness and adaptability of the proposed HAR approach across various circumstances, regardless of who is performing the activity.