Video-Based Dual-Task Four-Square Step Test for Fall Risk Assessment in Parkinson’s Disease

Abstract

Fall risk threatens independence in Parkinson’s disease (PD), particularly under combined cognitive–motor demands. This study proposes a video-based artificial intelligence (AI) framework to identify the most sensitive dual-task condition of the Four-Square Step Test (FSST) for fall-risk assessment. Thirty PD patients were recorded with a fixed RGB camera performing the FSST under single-task and three dual-task conditions: reciting days backward, holding a glass of water, and ball transfer. From each recording, silhouette masks and skeleton keypoints were extracted and fused into frame-level embeddings, fed into a multitask temporal convolutional network (TCN) jointly predicting task type and binary fall-risk. Videos were labeled high or low fall-risk using the clinician-measured 9.68 s threshold; model weights were initialized via transfer learning from 40 healthy adults. Clinically, completion times correlated with the Timed Up and Go test ( \(r = 0.50\) –0.66, \(p < 0.01\) ), and the reciting-days condition showed the strongest cognitive association (Montreal Cognitive Assessment, \(r = -0.48\) , \(p = 0.006\) ), supporting construct validity. From the AI model, the holding-a-glass condition achieved the highest fall-risk discrimination under transfer learning, as assessed by F1-score and AUROC, and the model-derived cutoff (9.65 s) matched the 9.68 s threshold. These findings suggest dual-task design enhances fall-risk sensitivity. Pending larger-scale validation, this framework may offer a scalable, video-based approach for fall-risk screening in PD.