Development of Falling Detection AI Model Using Neuro-Kinematic Multimodality in Parkinson

Parkinson's disease (PD) is a neurodegenerative disease that occurs in about 3% of the population over the age of 65. It causes a gradual loss of dopamine-producing nerve cells, resulting in a decline in motor function and an increased risk of falls, which significantly reduces the quality of life. Previous studies on fall prediction have analyzed neurological and kinematic factors separately. However, neurological factors alone are insufficient to fully reflect movement changes during a fall, and kinematic factors have limitations in explaining the underlying neurological mechanisms. In this study, we aimed to improve fall prediction performance by evaluating the complementary contributions of T1-weighted MRI and gait data through a multimodal approach. We analyzed 147 Parkinson's disease patients, including 25 fallers and 122 non-fallers. MRI data were segmented using U-NetR model to extract volumetric information for major brain structures, while gait data were analyzed using the GAITRite system to assess walking speed and stride length. The results using XGBoost showed that the MRI-only model achieved an AUROC of 0.78, the gait-only model 0.79, and the multimodal model 0.84. These results suggest that integrating brain structure and gait information can significantly improve fall prediction. This study emphasizes that multimodal approaches can provide important insights for early fall risk prediction and support clinical decision-making.