Wearable Sensor-Based Human Activity Recognition: Performance and Interpretability of Dynamic Neural Networks

Human Activity Recognition (HAR) using wearable sensor data is increasingly important in healthcare, rehabilitation, and smart monitoring. This study systematically compared three dynamic neural network architectures-Finite Impulse Response Neural Network (FIRNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU)-to examine their suitability and specificity for HAR tasks. A controlled experimental setup was applied, training 16,500 models across different delay lengths and hidden neuron counts. The investigation focused on classification accuracy, computational cost, and model interpretability. LSTM achieved the highest classification accuracy (98.76%), followed by GRU (97.33%) and FIRNN (95.74%), with FIRNN offering the lowest computational complexity. To improve model transparency, Layer-wise Relevance Propagation (LRP) was applied to both input and hidden layers. The results showed that gyroscope Y-axis data was consistently the most informative, while accelerometer Y-axis data was the least informative. LRP analysis also revealed that GRU distributed relevance more broadly across hidden units, while FIRNN relied more on a small subset. These findings highlight trade-offs between performance, complexity, and interpretability and provide practical guidance for applying explainable neural wearable sensor-based HAR.