Fall Detection Based on Recurrent Neural Networks and Accelerometer Data from Smartphones

• Published in: Applied sciences Vol. 15; no. 12; p. 6688
• Main Authors: Bartczak, Natalia, Glanowska, Marta, Kowalewicz, Karolina, Kunin, Maciej, Susik, Robert
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.06.2025
• Subjects: accelerometer; cellphone; Datasets; fall detection; Heart attack; Neural networks; recurrent neural networks; Smart phones; Smartphones; Support vector machines; UniMiB SHAR
• ISSN: 2076-3417; 2076-3417
• DOI: 10.3390/app15126688

An aging society increases the demand for solutions that enable quick reactions, such as calling for help in response to events that may threaten life or health. One of such events is a fall, which is a common cause (or consequence) of injuries among the elderly, that can lead to health problems or even death. Fall may be also a symptom of a serious health problem, such as a stroke or a heart attack. This study addresses the fall detection problem. We propose a fall detection solution based on accelerometer data from smartphone devices. The proposed model is based on a Recurrent Neural Network employing a Gated Recurrent Unit (GRU) layer. We compared the results with the state-of-the-art solutions available in the literature using the UniMiB SHAR dataset containing accelerometer data collected using smartphone devices. The dataset contains the validation dataset prepared for evaluation using the Leave-One-Subject-Out (LOSO-CV) and 5-Fold Cross-Validation (CV) strategies; consequently, we used them for evaluation. Our solution achieves the highest result for Leave-One-Subject-Out and a comparable result for the k-Fold Cross-Validation strategy, achieving 98.99% and 99.82% accuracy, respectively. We believe it has the potential for adoption in production devices, which could be helpful, for example, in nursing homes, improving the provision of assistance especially when combined into a multimodal system with other sensors. We also provide all the data and code used in our experiments publicly, allowing other researchers to reproduce our results.