An IOT-Driven Fall Detection System Using Bi-Directional LSTM for Safety in Elderly

Life expectancy is rising quickly along with the world’s population, especially in wealthier nations. Health care systems are facing serious challenges due to the growing proportion and total number of older persons in the population. As they age and become weaker, elderly people are falling more frequently, which makes it harder for them to remain stable all the time. Therefore, the goal of this research is to create a more intelligent and effective Deep Learning model for fall detection in older populations. In this study, a multi-modal dataset from barometer, magnetometer, accelerometer, and gyroscope motion signals we fused together to improve model generalizability and robustness. The data pre-processing involved data fusion, label processing, feature selection, and data transformation. We implemented three (3) different models which include Random Forest, Bi-LSTM and CNN-LSTM. The Bi-LSTM model performed best with an accuracy, precision, recall and F1-score of 97% for all the performance metrics while Random Forest and CNN-LSTM achieved 89% and 83% respectively for all the performance metrics.  Bi-LSTM model performance can be attributed to the adequate data preprocessing and its capability to learn and preprocess sequential data in both backward and forward directions as opposed to the Random Forest model with a very low adaptability to sequential data, while the  CNN-LSTM have the capability to learn from sequential data, its strength lies in image sequence dataset. This study can be adapted for the improvement in the field of fall detection.