A low-latency deep learning approach for human action recognition in medical internet of things applications

• Temporal-SpatialCNN: A lightweight framework enables low-latency (9.6 ms) and efficient (0.25 M parameters) action recognition on edge devices. • Novel Methodology: A specialized CNNBlock integrates four skeleton-derived features to optimize spatio-temporal learning. • Superior Performance: Validation across five benchmarks achieves state-of-the-art accuracy, reaching up to 99.66 % on the Florence-3D dataset. • Real-World Application: A real-time SOS gesture detection system confirms practical feasibility with minimal memory usage (0.94 MB). Real-time human action recognition (HAR) plays a vital role in healthcare monitoring, particularly for elderly care and assistive environments. However, existing HAR systems often struggle with high computational demands, large model sizes, and vulnerability to background noise, limiting their use on edge devices in the Internet of Medical Things (IoMT) settings. This study proposes Temporal-SpatialCNN, a lightweight framework built on a novel CNNBlock architecture that integrates convolution, batch normalization, and Spatial-Dropout to enhance both efficiency and generalization. The study systematically analyzes various layer arrangements within CNNBlock and identifies an optimal configuration that improves recognition performance while maintaining minimal computational overhead. The model incorporates diverse skeletal input modalities – joint coordinates, joint collection distances, slow motion, and velocity – to capture enriched spatio-temporal features. Extensive experiments on five well-known benchmark datasets validate the effectiveness of the proposed approach, achieving state-of-the-art accuracy (up to 99.66% on the Florence-3D dataset) with an inference time of 9.6 ms. To demonstrate real-world applicability, a real-time Save Our Soul (SOS) system was implemented on standard hardware, capable of detecting emergency gestures such as calls for assistance, thereby highlighting the model’s practical potential in real-time, resource-constrained healthcare scenarios.