Routine Blood Indicators Combined With Sleep Questionnaires for the Identification of Cognitive Impairment in Parkinson's Disease: A Cross-Sectional Study With a Leakage-Free Nested Modeling Framework.

LASSO feature selection identified 10 predictors of cognitive impairment in Parkinson's disease from 20 candidate variables.

• The 10 retained predictors were: education, disease duration, Hoehn-Yahr stage, platelet-to-lymphocyte ratio (PLR), monocyte-to-lymphocyte ratio (MLR), homocysteine, C-reactive protein (CRP), Pittsburgh Sleep Quality Index (PSQI), Epworth Sleepiness Scale (ESS), and REM Sleep Behavior Disorder Screening Questionnaire (RBDSQ).
• Feature selection used L1-penalized logistic regression with the lambda.1se criterion, confined strictly to training data (70%, n=242) to prevent data leakage.
• The cohort comprised 347 PD patients: 143 classified as PD-CI and 204 as PD-NCI based on MoCA < 26.

The full logistic regression model achieved an AUC of 0.943 in the untouched test set, significantly higher than the baseline clinical model.

• Full model AUC was 0.943 (95% CI: 0.896–0.990) in the held-out test set (n=105).
• Baseline clinical model AUC was 0.844; the incremental gain over baseline was statistically significant (DeLong p=0.009).
• Repeated 5×2 cross-validation yielded a more conservative AUC of 0.921 (SD=0.029).
• Calibration was acceptable: slope 1.18, intercept 0.03, test Brier score 0.086 (95% CI 0.057–0.120).

The baseline-plus-blood model and the full model showed comparable discriminative performance in the test set.

• Baseline-plus-blood model AUC was 0.937 and baseline-plus-sleep model AUC was 0.889 in the test set.
• The full model AUC (0.943) was comparable to baseline-plus-blood (0.937) but higher than baseline-plus-sleep (0.889) and baseline alone (0.844).
• These comparisons were assessed using DeLong confidence intervals and between-model DeLong tests.

Machine learning classifiers (random forest and gradient boosting) did not outperform logistic regression in discriminating cognitive impairment.

• Random forest achieved an AUC of 0.956 and gradient boosting achieved 0.944 in the test set.
• Neither RF nor GBM outperformed the full logistic regression model (AUC 0.943).
• Four nested logistic regression models were constructed alongside these two machine learning classifiers for comparison.

Model findings were stable across alternative MoCA cut-offs, supporting robustness of the outcome definition.

• AUC ranged from 0.909 to 0.931 across alternative MoCA cut-offs.
• Robustness was also examined with education-adjusted MoCA thresholds.
• The primary outcome definition used MoCA < 26 to classify PD-CI.

The study employed a strictly leakage-free nested modeling framework to prevent data leakage during model development.

• The cohort was divided into training (70%, n=242) and test (30%, n=105) sets by stratified sampling before any modeling steps.
• All feature selection and hyperparameter tuning were confined to the training data; the test set was scored only once.
• Generalization was additionally estimated by repeated 5×2 cross-validation.
• This design is described as a key methodological safeguard against optimistic performance estimates.

The authors explicitly state that the model should be regarded as a candidate screening aid requiring external validation before any clinical use.

• Limitations acknowledged include cross-sectional design, single-center recruitment, and only internal validation.
• The authors call for 'external, preferably multicenter and prospective, validation before any clinical use.'
• The study is described as quantifying 'the strength of association and the cross-sectional classification value' rather than establishing causal or predictive utility.