Development of a dynamics-enhanced facial algorithm to discriminate progressive supranuclear palsy (PSP) from...

Progressive supranuclear palsy (PSP) and Parkinson's disease (PD) share overlapping motor features, making early differential diagnosis challenging in clinical practice. Facial hypomimia is common in both conditions, but disease-specific differences in facial dynamics may provide discriminative diagnostic cues.

We recruited 62 participants (27 PSP and 35 PD) from a tertiary movement-disorders center. Facial videos were recorded during a standardized expression paradigm including smiling, anger, surprise, and rapid blinking. Dynamic geometric and velocity-based facial features were extracted using a 468-point MediaPipe facial landmark framework and normalized to physiologically meaningful reference distances. Machine-learning models-including support vector machine (SVM), random forest (RF), and gradient boosting (GB)-were trained using nested cross-validation. An independent 20% hold-out set was reserved for external-style validation. Model performance was evaluated using accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). Feature importance was assessed using F-scores, mutual information, RF importance, and SHAP analysis.

In cross-validation, the RF model achieved accuracy of 0.902, compared with 0.875 for SVM and 0.888 for the GB model. In the independent hold-out set, the RF classifier achieved an accuracy of 0.82 and an AUC of 0.77 (95% CI 0.63-0.89). Key features associated with group differences included higher mouth-opening velocity during surprise, reduced brow-eye distances, and greater forehead-wrinkle variability in PSP compared with PD.

Quantitative facial dynamics may provide interpretable digital biomarkers that can aid in differentiating PSP from PD. This facial-dynamics-based approach shows promise as a non-invasive adjunct for differential diagnosis, particularly in patients with overlapping early clinical presentations.