Multi-cohort machine learning identifies predictors of cognitive impairment in Parkinson’s disease
Title
Predicting Cognitive Decline in Parkinson’s Disease
One-Sentence Summary
Machine learning models trained on clinical data from three independent patient cohorts identified age at diagnosis and visuospatial ability as stable predictors of cognitive decline in Parkinson’s disease.
Overview
Cognitive impairment is a common non-motor symptom in Parkinson’s disease (PD), but its early prediction remains challenging. This study aimed to develop machine learning models to predict cognitive decline by integrating clinical data from three independent PD cohorts (LuxPARK, PPMI, and ICEBERG). The models were trained to predict two outcomes: mild cognitive impairment (PD-MCI), an objective measure, and subjective cognitive decline (SCD), a patient-reported measure. The multi-cohort models, which combined data from all three groups, demonstrated more stable performance than models trained on single cohorts, while maintaining competitive predictive accuracy. For instance, the multi-cohort model for classifying PD-MCI achieved an Area Under the Curve (AUC) of 0.67, and the model for SCD classification reached an AUC of 0.72. Key predictors consistently identified across cohorts included age at PD diagnosis and visuospatial ability. The analysis revealed that patients diagnosed at age 53 or older had a nearly 2.4-fold higher risk of developing PD-MCI.
Novelty
The study’s main contribution lies in its multi-cohort approach. While previous machine learning studies for PD cognitive impairment have typically relied on data from a single patient group, this research integrated data from three distinct international cohorts. This methodology enhances the generalizability and robustness of the predictive models. By training and validating the models across diverse populations, the findings are less susceptible to cohort-specific biases, such as differences in patient demographics or clinical assessment protocols. This cross-cohort strategy represents a significant step toward developing more universally applicable predictive tools for clinical use.
My Perspective
The divergence in key predictors for objective PD-MCI and subjective SCD is particularly insightful. The model for PD-MCI heavily weighted neurocognitive test scores like visuospatial performance, whereas the SCD model highlighted non-motor symptoms such as sleep disturbances and autonomic dysfunction, as well as male sex. This suggests that the biological processes underlying measurable cognitive decline may differ from the factors influencing a patient’s self-perceived cognitive difficulties. This distinction is crucial; it implies that effective clinical management may require a dual approach that addresses not only objective cognitive performance but also the broader spectrum of non-motor symptoms that shape a patient’s daily experience and quality of life.
Potential Clinical / Research Applications
Clinically, the identified predictors could help stratify patients at high risk for cognitive decline. Clinicians could use factors like older age at diagnosis and poor visuospatial test performance to identify individuals who may benefit most from early interventions, such as cognitive training or medication adjustments. In the long term, these predictive models could be integrated into digital health platforms for remote patient monitoring and screening. For research, this study validates the multi-cohort machine learning framework as a powerful method for identifying robust predictors, a strategy that could be applied to other PD symptoms or different neurodegenerative diseases. Furthermore, the distinct predictor sets for PD-MCI and SCD encourage future research into the different neurobiological mechanisms underlying objective versus subjective cognitive changes.