Deep Learning-Based Gait Event Prediction through a Single Waist-worn Wearable Sensor

Abstract

As the population of the elderly in the world continues its rapid surge, more attention must be drawn towards remote health monitoring (RHM), early diagnosis, and preventive interventions to sustain health care and eldercare. Elderly gait is a sensitive marker for their health status which enables detection of physical and mental impairments. This study aimed to advance the performance of elderly gait event detection using a single waist-worn IMU sensor through deep learning models. Four deep learning models MLP, CNN, LSTM, and GRU were trained and tested on data from the community-dwelling elderly. The accuracy was measured for six detection delay tolerance ranges. The GRU model achieved the best accuracy of 99.47% at a detection delay tolerance range of +/- 6TS (+/- 6ms) and 78.98% for a more precise range of +/- 1TS (+/- 1ms). The proposed method showed significant improvement over previously reported event detection results by achieving an MAE of 8.9ms and 7.6ms for the HS and TO events respectively. The RNN models attained better accuracy for the TO events as compared to HS events. Furthermore, an ablation study was also performed to observe the contribution of the subsets of the six IMU signals from the pelvis to the overall performance of the best model. The study showed that deep learning-based methods are much superior to previously reported methods for gait event detection through the pelvis sensor and it is expected to bring new attention towards practical waist-worn wearable devices for the health and condition monitoring of the elderly.