Interpretable machine learning models for identifying cognitive impairment in middle-aged and older adults with mild and severe insomnia: development, temporal validation, and clinical external validation.

Background: Cognitive impairment is an important health issue in middle-aged and older adults, and insomnia may be associated with increased cognitive vulnerability. However, models specifically designed to identify cognitive impairment in individuals with different severities of sleep-duration-defined insomnia remain limited. This study aimed to develop and validate interpretable machine learning models for current cognitive impairment identification in mild and severe insomnia subgroups. Methods: Data from CHARLS 2015 were used as the development cohort, CHARLS 2011 as the cross-wave temporal validation cohort, and clinical data from Gansu Provincial People's Hospital as the clinical external validation cohort. Participants with insomnia were stratified into mild and severe subgroups according to self-reported nighttime sleep duration. LASSO regression was used for feature selection, and candidate machine learning algorithms were compared for model selection. The selected LightGBM model was further evaluated using Bayesian optimization and optimized-threshold analysis. Model performance was assessed using AUROC, Brier score, calibration curves, decision curve analysis, and SHAP-based interpretability analysis. Results: The development, cross-wave temporal validation, and clinical external validation cohorts included 5,500, 4,231, and 500 participants, respectively. LightGBM showed the most balanced overall performance. In the mild insomnia subgroup, LightGBM achieved AUROCs of 0.772, 0.749, and 0.748 across the three cohorts; in the severe insomnia subgroup, the corresponding AUROCs were 0.763, 0.757, and 0.750. Bayesian optimization produced comparable external validation discrimination, while optimized-threshold analysis improved threshold-dependent classification performance. SHAP analysis suggested different feature contribution patterns across insomnia severity. Conclusion: LightGBM provided moderate and relatively stable performance for identifying current cognitive impairment risk in both insomnia subgroups. Combined with SHAP interpretation and online calculators, these models may support auxiliary screening, preliminary risk stratification, and referral prioritization for formal cognitive assessment, but should not be interpreted as standalone diagnostic tools or tools for predicting future incident cognitive impairment in sleep medicine settings.