Wearable gait analysis for differentiating progressive supranuclear palsy and Parkinson's disease: Clinically...

Early differentiation of progressive supranuclear palsy from Parkinson's disease is challenging due to overlapping motor features, particularly in early disease stages where clinical misclassification is common. Wearable inertial sensors provide high-resolution gait data that may reveal disease-specific and clinically relevant signatures.

To evaluate whether sensor-derived gait measures can distinguish progressive supranuclear palsy from Parkinson's disease using machine learning, and to compare performance between raw time-series signals and aggregate spatiotemporal variables, with emphasis on clinically interpretable gait biomarkers.

In this retrospective study, 34 participants with progressive supranuclear palsy and 410 with Parkinson's disease completed an instrumented Timed-Up-and-Go while wearing six synchronized sensors. Two input modalities were analyzed: raw accelerometer, gyroscope, and magnetometer signals sampled at 128 Hz; and 48 curated spatiotemporal features. Eight classifiers, including logistic regression, support vector machine, k-nearest neighbors, neural network, decision tree, random forest, extreme gradient boosting, and a baseline, were trained to classify progressive supranuclear palsy versus Parkinson's disease at the participant level using cross-validation procedures designed to prevent data leakage. The primary performance metric was macro-averaged F1-score. Shapley Additive Explanations were applied to aggregate-feature models to identify key discriminative gait features.

Raw time-series models performed strongly, with gradient boosting achieving the highest macro-F1 (0.87), followed by random forest (0.86) and support vector machine (0.83). Aggregate models performed comparably, with gradient boosting reaching 0.88, followed by random forest (0.87) and support vector machine (0.84). Shapley analyses identified increased double-support time and mediolateral sway, and reduced stride length, gait speed, and trunk range of motion as the most distinctive gait impairments characterizing PSP relative to PD.

Ensemble models reliably differentiated progressive supranuclear palsy from Parkinson's disease. Aggregate spatiotemporal models achieved performance comparable to raw signals while yielding clinically interpretable and physiologically meaningful gait biomarkers. These findings support the use of wearable sensor-based gait analysis combined with machine learning as a practical tool for improving differential diagnosis in movement disorder clinics.