Applying deep learning technology for automatic fall detection using mobile sensors. (February 2022)

• Record Type: Journal Article
• Authors: Wu, Xiaodan; Zheng, Yumeng; Chu, Chao-Hsien; Cheng, Lingyu; Kim, Jungyoon
• Is Part Of: Biomedical signal processing and control. Volume 72(2022)Part B
• Publication Date: 2022-02
• Subjects: Fall detection -- Mobile sensors -- Machine learning -- Deep learning -- Gated Recurrent Units (GRU); Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.103355
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Falls are one of the most common issues affecting the health of the elderly. Extracting effective features is critical to fall detection. Existing studies used automatic feature extraction with complex architectures. We proposed a novel Gated Recurrent Units (GRU) model. With the proposed GRU model we achieve outstanding classification results. Abstract: With improved quality of life, many countries are facing a serious aging problem. Falls, one of the most common issues affecting the health of the elderly, are likely to cause irreversible damage to them. Therefore, the problem of how to accurately identify falls has great research value. Extracting features that can effectively represent a fall is the key to detecting them. Conventional machine learning (ML) methods which extract features manually are tedious and time-consuming. In contrast, deep learning (DL) can autonomously learn features from input data; however, configuring its architecture is somewhat complex, and the training time is long. In this study, a novel DL model, the Gated Recurrent Units (GRU) architecture, is proposed to obtain high-level features for classification. We evaluate its relative performance against six popular ML-based classifiers and three DL architectures using two popular open-source datasets, collected using mobile sensors. Our results show that the proposed method outperformed other algorithms in nearly all of the five performance metrics we examined, for the datasets we tested. The accuracy of prediction reached 99.56% and 90.69%, and the F1 score reached 96.83% and 87.29%, respectively, showing good performance.