JMIR Publications

ABSTRACT

Background:

Fall detection is of great significance in safeguarding human health. By monitoring the motion data, fall detection systems (FDS) can monitor and promptly detect whether an individual has experienced a fall, triggering an alert to the fallen person or their family. Recently, wearable sensors-based FDS has become the mainstream of research. Different from threshold-based FDS and machine learning-based FDS, deep learning-based FDS, exemplified by Convolutional Neural Networks (CNN), can autonomously extract features from motion data. However, deep learning-based FDS still faces challenges, mainly in the insufficient attention to local information of sensor data and difficulty in understanding human posture and motion patterns highly related to fall behavior.

Objective:

This study aims to develop and validate a deep learning framework using acceleration and gyroscope data to accurately distinguish falls from activities of daily life (ADL).

Methods:

Based on three-axis acceleration and gyroscope data, we proposed a new deep learning architecture, the dual-stream CNN self-attention (DSCS) model. The model consists of three modules: feature extraction module, self-attention (SA) module and classification module. The proposed model was trained and tested on two public datasets: Sisfall and MobiFall. In addition, 10 participants were recruited to carry out practical validation of the DSCS model. A total of 1700 trials were performed to test the generalization ability of the model.

Results:

The fall detection accuracy of DSCS is 99.32% (sensitivity: 99.14%; specificity: 99.39%) and 99.65% (sensitivity: 100.0% ; specificity: 99.55%) on the test set of SisFall and MobiFall, respectively. In the ablation experiment, we compared the DSCS with state-of-the-art machine learning and deep learning models. On the SisFall dataset, DSCS achieved the best sensitivity, second-best accuracy; on the MobiFall dataset, DSCS achieved the best accuracy, sensitivity and specificity. Compared with the single-stream architecture (acceleration-only) and single-stream (gyroscope-only) architecture, DSCS could improve fall detection accuracy by 0.51% and 0.93% on SisFall, 0.35% and 1.76% on MobiFall. Compared with the architecture without SA module, DSCS could also improve fall detection accuracy by 1.02% on SisFall dataset and 1.41% on MobiFall dataset. In practical validation, the accuracy of the model was 96.41% (sensitivity: 95.12%; specificity: 97.55%).

Conclusions:

This study shows that DSCS model can effectively enhance the accuracy of fall detection on two public datasets, and performs well in practical validation.