Predicting Cognitive Impairment in Parkinson's Disease: A Machine Learning Approach Based on Clinical and Neuropsychological Data

Meili YangChuxin WangJinying ZhangYao XiaoYafang ChenZeming GuoJiayin WangJinzhong Huang^*

• The Second Affiliated Hospital of Fujian Medical University, Quanzhou, China

Abstract Background Cognitive impairment (CI) is a common and disabling non-motor symptom of Parkinson’s disease (PD), significantly reducing quality of life and increasing caregiver burden. Despite extensive research, early prediction of CI remains challenging owing to diverse clinical presentations, inconsistent treatment adherence, and inherent limitations in the sensitivity of conventional biomarkers and cognitive assessment tools. Methods and materials This retrospective cohort study included 514 patients with PD who had complete baseline data and at least 6 months of follow-up. The participants were randomly divided into training (n = 359) and test cohorts (n = 155). Demographic, clinical, biochemical, and neuropsychological data were obtained at baseline. CI was defined by Mini-Mental State Examination scores below education-adjusted thresholds and validated using the Montreal Cognitive Assessment (MoCA). Multiple machine learning (ML) models, including random forest (RF), logistic regression (LR), gradient boosting, CatBoost, and support vector machines, were developed and evaluated using the area under the receiver operating characteristic curve (AUROC), accuracy, recall, F1-score, calibration, and decision curve analysis. Feature importance analysis identified the key predictive variables. Results During follow-up, patients who developed CI were characterized by a significantly older age; longer disease duration; lower levels of albumin, hematocrit (HCT), and blood lipids; and higher prevalence of hypertension. Feature selection identified age, platelet count, time from diagnosis to baseline visit, apolipoprotein B level, and HCT level as the predictors. The RF model demonstrated the best overall performance, with AUROC of 0.846, accuracy of 0.75, and F1-score of 0.775, followed by CatBoost and LR. Calibration and decision curve analyses confirmed stable probability estimation and superior clinical utility of RF compared with those of “treat all” or “treat none” strategies. The MoCA score was further used to verify the model’s stability. Conclusions ML models that integrate multimodal clinical and neuropsychological data demonstrate high accuracy in predicting CI in Parkinson’s disease, with RF emerging as the most reliable approach. This framework provides a practical tool for early risk stratification, potentially enabling timely interventions and individualized management to reduce the burden of cognitive decline in PD.