Single-Wearable Sensor-Based Elderly Fall and Activity Recognition Using CNN-biLSTM: A Comparative Study with CNN, LSTM, and CNN-LSTM

The growing elderly population highlights the need to support their independence and manage risks, with falls being a leading cause of death among them. Many studies use wearable sensors to recognize falls and daily activities in the elderly by attaching them to several parts of the body simultaneously. This causes discomfort and limits the movement of the elderly. This study proposes recognition of falls and activities in the elderly using sensors attached to the abdomen with a deep learning approach, namely CNN-biLSTM. To make an objective comparison, we also built CNN, LSTM, and CNN-LSTM model with the same environment. The experiment was conducted using the SisFall public dataset taken in a lab environment. To enrich the diversity of the dataset, we added it with a dataset from our own collection. Our dataset was collected in the elderly's home environment using ESP32 and a QMI8658 Inertial Measurement Unit (IMU). The spatial characteristics of the sensor can be effectively captured using CNN. Meanwhile, biLSTM is used to handle temporal motion patterns in two directions: forward and backward. The combination of the two is able to produce better classification performance. The results showed that the combination of CNN-biLSTM models provided the best accuracy, recall, precision and F1-Score compared to CNN, LSTM, and CNN-LSTM models. The accuracy of the CNN-biLSTM model was 96.89%, CNN 94.07%, LSTM 87.46%, and CNN-LSTM 95.41%. CNN-biLSTM is the best model in recognizing falls, lying, sitting, and standing activities based on F1-Score.