Parkinson’s disease assessment with selfie video via feature engineering and explainable AI

Abstract

Parkinson’s disease (PD) is a neurodegenerative disorder fundamentally characterized by motor impairments and diminished facial expressivity. This study investigates the efficacy of feature engineering approaches applied to facial video data captured via smartphone frontal cameras to predict PD presence, disease severity, and motor function scores. Within a cohort of 40 participants, a combination of three predictive models, four facial movement scenarios, and five frame selection strategies was rigorously evaluated. Regarding PD classification, the XGBoost model achieved maximum performance using the smiling video scenario with frame skipping combined with PCA retaining 95% of the variance, yielding an accuracy of 0.77, a precision of 0.75, a recall of 0.86, and an F1-score of 0.79. For Hoehn & Yahr severity estimation, the Random Forest model implemented with the smiling video, frame skipping, and no PCA attained the best result, with the lowest RMSE of 1.30 and a MAPE of 38.36%. Furthermore, in The Movement Disorder Society-sponsored revision of the Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) Part III motor score estimation, the XGBoost model outperformed other configurations using the smiling video with 3-frame averaging and PCA retaining 95% of the variance, achieving an RMSE of 19.34 and a MAPE of 45.33%. Crucially, the Explainable AI analysis identified that lip and mouth corner movements were the most critical features for both PD classification and Hoehn & Yahr estimation. In contrast, eyebrow and eye movements were the dominant features for MDS-UPDRS Part III motor score estimation. These findings indicate that facial dynamics carry information relevant to PD assessment and support further development of non-invasive, accessible screening tools, although the accuracy achieved here is not yet sufficient to substitute for clinical evaluation.