An XGBoost machine learning model using just six readily available clinical and demographic variables, family history of diabetes, education level, marital status, chronic disease burden, chronic pain, and depression, can predict sleep disorder risk in diabetic patients with 85% accuracy, researchers report in Scientific Reports.
The study, drawing on data from the China Health and Retirement Longitudinal Study (CHARLS), screened over 60,000 elderly individuals and identified 1,276 diabetic patients, of whom 499 (39.1%) had sleep disorders.
What they found
The research team led by Maoqin Tian tested five machine learning classifiers, logistic regression, decision tree, XGBoost, support vector machine, and LightGBM, on the CHARLS dataset. Feature selection was performed using single-factor correlation analysis followed by LASSO regression, which narrowed the candidate predictors to six:
- Family history of diabetes
- Education level
- Marital status
- Number of chronic diseases
- Chronic pain
- Depression
XGBoost outperformed all other models, achieving an area under the receiver operating characteristic curve (AUC) of 0.850. Calibration curves confirmed good fit, and decision curve analysis showed excellent net benefit, meaning the model could be useful in clinical decision-making without incurring excessive false positives.
Why it matters
Sleep disorders are highly prevalent in patients with diabetes but are frequently undiagnosed. Polysomnography is expensive and inaccessible for many, while self-report screening tools have limited accuracy in this population.
A predictive model that relies on six easily collectable variables, no lab tests, no wearable devices, no questionnaires beyond routine history-taking, could be integrated into primary care workflows or electronic health record systems to flag diabetic patients who would benefit from formal sleep evaluation. If externally validated, the tool could help bridge the screening gap in settings where sleep medicine resources are scarce.
The CHARLS dataset is representative of China’s aging population, making the findings particularly relevant for countries facing rapid demographic aging and rising diabetes prevalence.
Limits
The model was developed and tested on a single Chinese cohort and has not been externally validated in other populations or healthcare systems. The study used a retrospective design based on self-reported sleep disturbance, not objective sleep measures such as actigraphy or polysomnography. The relatively modest sample of 1,276 diabetic patients limits statistical power for subgroup analyses. Prospective validation studies with objective sleep outcomes are needed before clinical deployment.
Bottom line
An XGBoost model using six basic demographic and clinical variables predicted sleep disorder risk in diabetic patients with an AUC of 0.85. If validated externally, it could provide a low-cost screening tool for identifying which diabetic patients need formal sleep evaluation.
Source
Maoqin Tian et al. Machine learning-based predictive model for sleep disorders in diabetic patients: data analysis from CHARLS. Scientific Reports (2026). DOI: 10.1038/s41598-026-53312-x. PMID: 42315864.