PSG-Based Prediction of Acupuncture Efficacy in Insomnia

Insomnia is a chronic health condition that significantly impairs patients’ quality of life and mental well-being. Although acupuncture is widely used in the treatment of insomnia due to its safety and efficacy, significant inter-individual variability in treatment response and the absence of a pre-treatment predictive mechanism limit the development of precise medication. To address this issue, a predictive framework integrating feature engineering and multi-model ensemble learning was constructed based on pre-treatment polysomnographic parameters. To handle the small sample size and class imbalance, adaptive synthetic sampling was applied to expand the training set, and recursive feature elimination with cross-validation was used to select the optimal feature subset. Five classification models—support vector machine, random forest, logistic regression, XGBoost, and artificial neural network—were compared using stratified 3-fold and leave-one-out cross-validation. Results indicated that support vector machine achieved an accuracy of 80.00% and a weighted F1-score of 83.75% under stratified cross-validation, but its performance declined in leave-one-out validation. In contrast, random forest and XGBoost demonstrated consistent stability and robustness across both validation methods, with random forest achieving the highest leave-one-out cross-validation accuracy of 90.00% and an F1-score of 85.26%. SHAP analysis revealed that "sleep latency" and "deep sleep" were consistently identified as important features across multiple models, demonstrating strong physiological interpretability and generalizability. A simplified prediction rule based on these key features achieved a sensitivity of 83% and specificity of 71% in an independent test set, enabling effective identification of potential high-responders. A stratification strategy based on prediction results was further proposed, and prospective validation confirmed its efficacy in significantly improving the treatment response rate among initial low-responders.