Parkinson’s disease (PD) shows wide heterogeneity in symptom onset, severity, and progression, which complicates predicting individual clinical trajectories. The authors applied an interpretable machine learning approach to identify predictors for two clinically relevant endpoints in early PD: progression to mild cognitive impairment (MCI) and development of postural instability (PI). The goal was to determine whether a minimal clinico-demographic feature set could achieve comparable predictive accuracy to an extended multimodal dataset that includes imaging and fluid biomarkers, while preserving model interpretability to support personalized clinical decision-making.
The analysis used data from two independent cohorts of early PD patients. Data from the PPMI study were downloaded on January 27, 2023 from the PPMI database. Additional cohort data (referred to as the TuePaC cohort) are available upon request from the corresponding author. The authors made the code for their analyses publicly accessible at the GitHub repository: https://github.com/meDIC-UKT/LOC-PD.
The investigators used Explainable Boosting Machines (EBM)—an interpretable machine learning method—to build predictive models for progression to MCI and PI. EBMs are designed to provide both competitive predictive performance and transparent feature-level explanations, allowing identification of the most influential predictors in a way that can be communicated back to clinicians and patients.
Two types of input feature sets were compared:
A minimal clinico-demographic dataset, comprising basic demographic and clinical measures.
An extended multimodal dataset that incorporated additional domain information, including volumetric brain imaging and fluid biomarker data, alongside clinical and demographic variables.
The authors report that EBM models trained on the minimal dataset achieved predictive performance comparable to models trained on the extended multimodal dataset, indicating that a limited set of readily available clinical variables may suffice for risk stratification in early PD when using an interpretable ML approach.
Using the minimal feature set, the two most important predictors identified for both progression to MCI and to PI were baseline age and the Montreal Cognitive Assessment (MoCA) total score. When the extended multimodal dataset was used, additional predictive contributors emerged, including motor measures, other cognitive features, and imaging-derived variables. The source reports these additional features were identified by the extended models but does not provide a full ranked list or effect sizes in the abstract.
The study emphasizes that EBMs delivered comparable predictive accuracy using minimal inputs versus richer multimodal inputs, while maintaining interpretability. This balance allows clinicians to see which predictors drive individual risk estimates. Specific performance metrics (for example, AUC, sensitivity, specificity) and detailed comparative statistics between minimal and extended models were not reported in the abstract and therefore are not available from the source text provided here.
The authors propose that interpretable ML predictions of cognitive and motor endpoints in early PD could bring personalized clinical decision-making closer to practice, by identifying patients at higher risk of early MCI or PI. The finding that baseline age and MoCA total score are dominant predictors suggests that routine clinical assessment may capture substantial prognostic signal. However, details about absolute predictive performance, calibration, external validation beyond the two cohorts, and potential limitations such as sample representativeness or missing-data handling were not reported in the source abstract and would require consultation of the full manuscript or supplementary materials for clarification.
The study received approval from the Ethics Committee of the University of Tuebingen, and informed consent procedures were followed. Funding and competing interest statements are detailed in the source: several authors report funding from and relationships with research foundations, industry, and other organizations; individual disclosures are enumerated in the original report. The PPMI data used are publicly available via the PPMI database, TuePaC cohort data are available on request from the corresponding author, and analysis code is hosted at https://github.com/meDIC-UKT/LOC-PD.
In two independent early-PD cohorts, interpretable machine learning using Explainable Boosting Machines identified baseline age and MoCA total score as the primary predictors of progression to MCI and postural instability when using a minimal clinico-demographic feature set. Extended multimodal models added motor, cognitive, and imaging predictors but did not markedly outperform the minimal models according to the source abstract. The work supports the potential role of explainable ML for individualized prognostication in early PD and highlights accessible clinical measures as core prognostic indicators. For full methodological and statistical details, consult the full preprint and associated supplementary material.