A Robust Machine Learning-Based Approach for Detecting Falls Using Waist-Worn Accelerometer Data

Falls are a leading cause of injury and hospitalization among older adults, necessitating early and dependable detection for timely medical intervention. This work introduces a fall detection framework built upon motion data obtained from a waist-mounted triaxial accelerometer. The system distinguishes fall events from typical daily activities by applying traditional machine learning techniques. The dataset includes both artificially induced fall scenarios and routine movements to enable realistic system validation. From the time-domain signals, significant motion descriptors are extracted and used to train several classifiers—including Random Forest, Support Vector Machine (SVM), and k-Nearest Neighbors (k-NN)—under a 5 fold cross-validation scheme. Of all evaluated models, Random Forest yielded superior results, achieving 96.3% classification accuracy and an F1-score of 95.8%. The outcomes affirm that an accelerometer-only solution, coupled with efficient machine learning models, can offer a reliable, real-time fall detection mechanism suitable for integration into wearable healthcare devices.