Machine learning-based prediction of multiple system atrophy and progressive supranuclear palsy using clinical and neuropsychological scores

Differentiating between multiple system atrophy (MSA) and progressive supranuclear palsy (PSP) remains clinically challenging owing to the significant overlap in symptoms. Herein, we evaluated whether machine learning assisted feature selection combined with logistic regression could distinguish MSA from PSP using clinical and neuropsychological data. This was a prospective single-center study with ≥3-year diagnostic follow-up (total n = 64; MSA n = 45, PSP n = 19). The descriptors used were the observed data. For modeling, continuous predictors were mean-imputed, and categorical predictors were assigned an explicit “Missing” level (outcome unaltered). An exploratory Support Vector Machine (SVM, radial) provided variable rankings (excluding the PSP Rating Scale to avoid incorporation bias). The primary classifier was an elastic-net penalized logistic regression with 11 prespecified predictors (Montreal Cognitive Assessment [MoCA], age, Scales for Outcomes in Parkinson's Disease-Autonomic Questionnaire [SCOPA-AUT] sexual status [men], SCOPA-AUT total, SCOPA-AUT urinary status, animal fluency, geriatric depression scale [GDS], Goldenberg imitation total, disease duration, Goldenberg finger imitation, and PSP Rating Scale). Hyperparameters were tuned by repeated stratified 5-fold cross-validation (20 repeats), and the performance was estimated from patient-level out-of-fold predictions. Participants with PSP were older than those with MSA (68.26 ± 8.74 vs 62.71 ± 8.18 years; P = 0.018). The exploratory ranking highlighted the cognitive and autonomic measures (MoCA, animal fluency, SCOPA-AUT sexual [men], Goldenberg imitation, age) among the most informative features. The elastic-net model achieved an area under the curve (AUC) of 0.80 (95% confidence interval [CI] 0.67–0.93) on out-of-fold predictions at Youden's J (threshold about 0.70), a sensitivity of 0.82, and specificity of 0.79. In comparison, the unpenalized logistic model showed an apparent AUC of 0.89 (95% CI 0.79–0.99). An interpretable penalized logistic model using routine measures distinguished MSA from PSP with balanced sensitivity and specificity in this small cohort. These findings are preliminary and warrant external validation, calibration assessment, and decision curve analysis prior to clinical application.