Insulin Resistance Surrogates and Cognitive Impairment in Parkinson’s Disease: A Cross-Sectional Study with Interpretable Machine Learning

<b>Background</b>: Insulin resistance (IR) has emerged as a key player in the pathogenesis of cognitive impairment in Parkinson's disease (PD). This study aims to systematically compare glucolipotoxicity-based (TyG, AIP) versus adiposity-driven (TyG-BMI, METS-IR) IR indices for their associations with PD dementia and to develop a clinically applicable nomogram using an interpretable machine learning framework. <b>Methods</b>: This cross-sectional study analyzed 251 PD patients: 42 with normal cognition, 160 with mild cognitive impairment (PD-MCI) and 49 with dementia (PDD). Logistic and linear regression examined associations between IR indices and cognitive impairment across different domains. Six machine learning models were compared for dementia classification, with the optimal model interpreted using SHapley Additive exPlanations (SHAP) to construct a nomogram. <b>Results</b>: Each standard deviation increase in TyG and AIP was linked to 79% (OR 1.79, 95%CI 1.04-3.07) and 75% (OR 1.75, 95%CI 1.05-2.91) higher risk of PDD, respectively, but not PD-MCI. In contrast, TyG-BMI and METS-IR showed no significant associations with either condition. TyG showed linear negative correlations with memory and orientation, and inverted U-shaped associations with visuospatial function and attention. AIP exhibited linear negative correlation with memory. The logistic regression model achieved the highest performance (AUC of 0.759) among six machine learning models. Crucially, SHAP analysis visually quantified TyG as a top modifiable predictor, facilitating the construction of an interpretable clinical nomogram. <b>Conclusions</b>: Glucolipotoxicity-based indices (TyG, AIP), unlike BMI-dependent markers (TyG-BMI, METS-IR), are robustly linked to PD dementia through domain-specific linear or nonlinear patterns. This suggests metabolic dysregulation predicts risk independent of weight loss. Furthermore, integrating SHAP-based interpretability transforms complex algorithms into a transparent, actionable tool for early risk stratification.