Towards Safer Aging: A Hybrid KNN Model for Pre-Impact Fall Detection Enhanced by Class Balancing

Abstract

Falls among the elderly are a significant public health issue, often resulting in serious injuries or even fatalities. Detecting falls before they happen (pre-impact) can enable timely interventions and help minimize the severity of injuries. This research focuses on creating an effective pre-impact fall detection system for elderly individuals by utilizing class balancing techniques and a hybrid k-nearest neighbors (KNN) approach. To address the challenges posed by imbalanced datasets, methods like SMOTE, ADASYN, and RUS were employed, while the hybrid KNN model was used for accurate classification. The system was tested using the KFall dataset, which includes sensor data from accelerometers and gyroscopes. The hybrid KNN model demonstrated an impressive 99.8% accuracy in detecting pre-impact falls, surpassing traditional KNN and other standard methods. Moreover, the use of class balancing significantly enhanced the detection of minority classes. This study highlights the potential of combining class balancing techniques with hybrid KNN models for reliable pre-impact fall detection. These findings could pave the way for developing wearable devices aimed at preventing falls and improving safety for elderly individuals.