Fusion of Dual Sensor Features for Fall Risk Assessment with Improved Attention Mechanism

Abstract

Nowadays, falls are one of the important causes of accidental death in older people. Assessment of fall risk can help to protect older adults in a timely manner. Current studies tend to use a single type of sensor, which always suffers from insufficient robustness, and the accuracy of the risk assessment model is low. In this study, we proposed a Convolutional Neural Network (CNN)-Bi Long Short-Term Memory (LSTM) fall risk assessment model based on the fusion of multi-sensor information with improved efficient channel attention (ISA-ECA-CNN-BiLSTM). Firstly, we construct a hybrid network consisting of a LSTM network and a CNN, which can capture the features hidden in asynchronous gait data sequences very well. An improved efficient Channel Attention Mechanism was also incorporated to make the model more attentive to the global features of the gait. Since the features extracted from the plantar pressure distribution signal and the IMU signal do not contribute to the fall risk assessment to the same extent, an adaptive weighted feature fusion method was introduced to enhance the influence of important features on the assessment results while weakening the influence of unimportant features on the assessment results. The improved method has higher sensitivity, specificity, and accuracy compared to the direct cascade method. The experimental results show that the accuracy, precision, sensitivity, and F1-score of the ISA-ECA-CNN-BiLSTM model proposed in this study were 98.4%, 99.1%, 98.8%, and 98.9%, respectively, which are higher than other classification models and can effectively extract gait features, thus improving the accuracy of fall risk recognition.