Interpretable machine learning for predicting moderate-to-severe insomnia risk in Chinese adults with autism spectrum disorder.

Objective: To develop a risk prediction model for moderate-to-severe insomnia among adults with autism spectrum disorder (ASD) and to identify key predictive features using interpretable machine learning methods, to support exploratory risk estimation and stratification. Methods: This study used data from the 2024 Psychological and Behavioral Survey Database of Adults with ASD and included 976 adults with ASD in the final analysis. Moderate-to-severe insomnia was defined as an Insomnia Severity Index (ISI) score ≥ 15. To avoid outcome leakage, all ISI items and the ISI total score were excluded from the predictor set, and sleep-related items were also removed from the depression and anxiety symptom scores. The dataset was first divided into training and test sets at a ratio of 7:3 using stratified random sampling. Candidate predictors were subsequently selected using least absolute shrinkage and selection operator regression based exclusively on the training set, and multicollinearity was assessed using the variance inflation factor (VIF). For models requiring feature scaling, standardization parameters were estimated from the training data only and subsequently applied unchanged to the test data. Nine machine learning models were developed using the training set and evaluated in the independent test set. The final model was interpreted using SHAP, and an online prediction tool was developed based on important variables. Results: Among the 976 adults with ASD, 265 participants (27.2%) were classified as having moderate-to-severe insomnia. After LASSO selection using the training set, 18 predictors were retained, and no substantial multicollinearity was observed. In the held-out test set, logistic regression achieved a ROC AUC of 0.7262 and a PR AUC of 0.4624, compared with a no-skill precision-recall baseline of approximately 0.273. At the operating threshold used for classification, sensitivity was 0.4875, specificity was 0.8216, and accuracy was 0.7304, which was similar to the no-information rate of 0.7270. SHAP analysis showed that the total score of the non-sleep items of the GAD-9, the total score of the non-sleep items of the PHQ-8, anxiety diagnosis, proportion of screen time spent watching short videos, age, and sex were among the predictors with higher contributions. Conclusion: This study developed an interpretable prediction model for moderate-to-severe insomnia risk among adults with ASD. Anxiety- and depression-related symptoms were the main predictive features, while the proportion of screen time spent watching short videos also provided additional predictive value. The online tool developed based on key variables may provide an exploratory approach to individualized risk estimation and stratification, but further validation and threshold optimization are required before it can be considered for screening applications.