Fig. 5: Model performance in discriminating the medication effect on gait impairments.

a The accuracies in discriminating the patient outcomes by three expert clinicians, a non-expert clinician, the AI model, and the significant changes in the traditional clinical motion markers. For the sub-UPDRS score assessment, three experts performed three rounds of independent assessments of the patient outcomes with different sets of information: 1) solely gait videos, 2) gait videos accompanied by medication status and disease details (Table 1), 3) gait videos with medication status and disease details, supplemented by values of traditional motion markers measured in both off and on medication states (Fig. 4). b The Spearman correlation coefficients (ρ) between the patient outcomes discriminated by clinicians, AI model, and the clinical motion markers, with thicker chords indicating stronger correlations. c Discrimination accuracies of the three experts, the non-expert clinician, and the AI model for different cohorts: 1) 7 patients with changed UPDRS scores after medication (SwC cohort), 2) 12 patients without changed UPDRS scores after medication (SwoC cohort). Note that for assessing the medication’s effect on gait impairments in the SwC cohort, the non-expert clinician used a granular three-level sub-score rating approach. The three experts' ratings were still derived from changes in the standard UPDRS scale.