A Lightweight MLP-Based Fall Detection System for Wearable Devices

Authors

Lin Yang, School of Engineering, Computer and Mathematical Sciences, Auckland University of Technology, Auckland, New ZealandFollow
W.K. Darsha Jayamini, School of Engineering, Computer and Mathematical Sciences, Auckland University of Technology, New ZealandFollow
Farhaan Mirza, School of Engineering, Computer and Mathematical Sciences, Auckland University of Technology, New ZealandFollow

Abstract

Fall detection systems face persistent challenges with high false alarm rates, undermining user trust in real-world deployments. This study investigates an optimized lightweight Multilayer Perceptron (MLP) that integrates class weighting and regularization strategies to address data imbalance on resource-constrained wearable devices. The model was evaluated on the SisFall public dataset and a self-collected validation dataset. Comparative analysis against Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) shows the optimized MLP achieves .57 fall recall (5.7× improvement over baseline) with .12s inference time (142× faster than CNN methods), satisfying computational efficiency requirements. External validation demonstrates high fall sensitivity (recall = 1.) but limited precision (.29) and overall accuracy (.5), indicating generalization challenges for deployment. These results underscore the critical need for domain adaptation strategies and diverse training data to bridge the gap between laboratory performance and real-world applicability.

Recommended Citation

Yang, Lin; Jayamini, W.K. Darsha; and Mirza, Farhaan, "A Lightweight MLP-Based Fall Detection System for Wearable Devices" (2025). ACIS 2025 Proceedings. 111.
https://aisel.aisnet.org/acis2025/111