Hybrid CNN-LSTM Deep Learning Framework for Robust Real-Time Human Activity Recognition in Wearable Healthcare Monitoring Systems

Abstract

Human activity recognition using wearable sensors plays a critical role in healthcare monitoring systems, yet existing approaches often struggle to simultaneously capture local spatial patterns and long-term temporal dependencies from noisy sensor data. In this paper, we propose a hybrid deep learning framework that sequentially integrates a convolutional neural network module for spatial feature extraction with a long short-term memory module for temporal dynamics modeling. The system first applies low-pass filtering and min-max normalization to preprocess raw tri-axial accelerometer and gyroscope signals; the continuous data stream is then segmented into fixed-size sliding windows. Convolutional layers with learnable filters detect invariant spatial patterns such as peaks and frequency components from the input time series, and max-pooling layers subsequently reduce feature dimensionality while preserving salient information. The resulting feature maps are reshaped and fed into the LSTM component, whose gating mechanisms effectively capture long-range temporal dependencies and mitigate the vanishing gradient problem. The hidden states output by the LSTM are passed through fully connected dense layers, and a softmax function generates a probability distribution over predefined activity classes. The entire architecture is optimized end-to-end using the Adam optimizer to minimize categorical cross-entropy loss. The proposed method is particularly suitable for wearable healthcare applications where both accuracy and real-time responsiveness are demanded. Furthermore, the proposed method generalizes well across individuals and sensor configurations, as the learned features are invariant to noise and sensor displacement