Journal of Intelligence and Information Systems · Mar 31, 2026
Impact Analysis of Short-Term Predictive Factors for Elderly Falls Using a Decision Tree
By Cheon S, Jang Y, Kim C et al.
As South Korea enters a super-aged society, falls among older adults are drawing attention as a major factor that increases physical, psychological, and economic burdens. This study aims to predict fall risk among older adults using video-based body pose data and to identify the key variables affecting falls. Addressing the limitation of previous studies that focused on ‘detection’ to identify the occurrence of falls, this study presents the feasibility of designing a ‘short-term prediction’ model that captures warning signs immediately before a fall. Key joint coordinates were extracted using GMDCSA24, a publicly available fall video dataset, and variables such as joint angles, angular velocity, angular acceleration, body center displacement, and movement pattern ratios were generated based on these coordinates. The data were standardized to 30FPS, and noise was removed through coordinate normalization and Kalman Smoothing before the data were converted into model inputs. Decision Tree was applied for fall classification and variable importance analysis. Performance evaluation results (F1-score, Accuracy, AUC) showed that body angles in older adults had the greatest influence on fall prediction, and the model that considered both body angles and variables related to body balance (Exp7) showed the highest predictive performance. These results suggest that fall risk among older adults can be effectively predicted using noncontact video analysis alone, and this approach is expected to contribute to the future development of technologies for predicting and responding to fall risk among older adults in real time as a practical alternative that addresses the limitations of sensor-based approaches.