Attention Based Multimodal Sensor Fusion for Fall Detection Using Wearable Time-Series Data

Abstract

Injuries and disabilities from falls are one of the leading causes of death, primarily in the world’s aging population, and is an even more severe problem within developing countries with fewer options for ongoing and consistent health care. The aim of study was to use a quantitative research design to analyze the effectiveness of attention-based multimodal sensor fusion techniques on fall detection involving wearable sensors within Pakistan. The study included a sample of 150 participants who were simulated performing both falls and other daily living activities while wearing wearable technologies with accelerometer and gyroscope sensors. Data collected were analyzed using existing attention-based multimodal fusion techniques on the wearable time series datasets, comparing the results against single sensor techniques, threshold-based techniques, and non-attention based deep learning techniques. Using only the accelerometer and gyroscope, each method achieved good accuracies, approximately 85.2% and 82.4% respectively. However, the results from the attention-based fusion approach for the various modalities showed significantly improved results for detecting falls with an accuracy of 92.8% (F1-score = 90.8% and false positive rate = 7.2%). The improvement in accuracy compared to the threshold-based approaches (p < 0.001) and non-attentional deep learning methods (p = 0.004) was statistically significant. Additionally, the feasibility analyses provided evidence that the system has strong real-time processing capabilities (M = 4.21) and feasibility for overall real-world deployment (M = 4.12). It is concluded that attention-based multimodal sensor fusion can increase both the accuracy and practicality of wearable fall detection systems making them suitable for implementation in low-resource health systems like that of Pakistan.