# A Feature Engineering Method for Smartphone-Based Fall Detection

## TL;DR

This paper introduces a smartphone-based fall detection method using machine learning that achieves high accuracy and transparency across multiple datasets.

## Contribution

A novel feature engineering approach for fall detection using KNN and SVM with SHAP interpretability analysis across diverse datasets.

## Key findings

- The method achieved 98.45% average accuracy on the UniMiB SHAR dataset using LOSO cross-validation.
- It reached 99.89% peak accuracy on MobiAct with KNN and 96.41% in cross-dataset validation.
- SHAP analysis provided insights into the most influential features for fall detection.

## Abstract

A fall is defined as an event in which a person inadvertently comes to rest on the ground, floor, or another lower level. It is the second leading cause of unintentional death worldwide, with the elderly population (aged 65 and above) at the highest risk. In addition to preventing falls, timely and accurate detection is crucial to enable effective treatment and reduce potential injury. In this work, we propose a smartphone-based method for fall detection, employing K-Nearest Neighbors (KNN) and Support Vector Machine (SVM) classifiers to predict fall events from accelerometer data. We evaluated the proposed method on two simulated datasets (UniMiB SHAR and MobiAct) and one real-world fall dataset (FARSEEING), performing both same-dataset and cross-dataset evaluations. In same-dataset evaluation on UniMiB SHAR, the method achieved an average accuracy of 98.45% in Leave-One-Subject-Out (LOSO) cross-validation. On MobiAct, it achieved a peak accuracy of 99.89% using KNN. In cross-dataset validation on MobiAct, the highest accuracy reached 96.41%, while on FARSEEING, the method achieved 95.35% sensitivity and 98.12% specificity. SHAP-based interpretability analysis was further conducted to identify the most influential features and provide insights into the model’s decision-making process. These results demonstrate the high effectiveness, robustness, and transparency of the proposed approach in detecting falls across different datasets and scenarios.