Comparative Analysis of AI-driven Feature Selection and Machine Learning Framework for Efficient Radar Fall Detection

Radar sensors are increasingly viable for fall detection, yet optimizing feature selection and machine learning (ML) model efficiency remains essential for practical deployment. While the literature extensively explores complex AI models for feature extraction and fall detection, a notable gap exists in optimizing the feature selection process for radar data to enhance computational efficiency. This paper addresses this gap by presenting a comprehensive comparative evaluation of AI-driven feature selection techniques -Variance Threshold (VT), Spearman Correlation, Recursive Feature Elimination (RFE) and embedded methods - applied across various ML classifiers to balance accuracy and processing speed for fall detection. Using FMCW radar data from 400 fall and 500 non-fall samples collected from volunteers aged 26 to 78, results demonstrate that tree-based classifiers, particularly Extra Trees with RFE, provide an optimal balance, achieving a 0.08 ms runtime with 98% accuracy, ideally suited for real-time precision-critical fall detection. Additionally, Random Forest and Gradient Boosting with VT method achieved comparable accuracy of 97.22% with runtimes of 0.12 and 0.17 ms, respectively. This work presents a scalable framework for fall detection with insights into the strengths and limitations of various feature selection strategies, advancing real-time radar-based fall detection and embedded system design in healthcare applications.