Machine Learning-Based Prediction of Sleep Quality in Patients with Multiple Sclerosis

Background/Objectives: Sleep disturbances are common but frequently underrecognized in multiple sclerosis (MS), independently predicting reduced quality of life and worsening fatigue, cognitive impairment, and depression. While pain, nocturia, fatigue, and mood symptoms are established contributors, cardiometabolic and inflammatory factors remain largely unexplored despite their known links to poor sleep in other populations. This study used interpretable machine learning to determine the relative contribution of disease-related, cardiometabolic, and inflammatory parameters to sleep quality, assessed by the PSQI, in patients with MS. Methods: This cross-sectional, observational, single-center cohort study enrolled adult patients with MS. Sleep quality (PSQI), daytime sleepiness (ESS), and restless legs syndrome severity (IRLS) were assessed alongside clinical, anthropometric, hemodynamic, and laboratory parameters, including inflammatory and metabolic indices. Three regression models, Support Vector Regression, Random Forest, and XGBoost, were trained on 28 predictors to predict PSQI global scores and evaluated on a independent test set, with feature contributions examined using SHAP analysis. Results: A total of 173 patients with MS were included (mean age 39.66 ± 11.86 years, 69.9% female, 90.2% RRMS), with 48.0% classified as poor sleepers (PSQI > 5) and a mean PSQI score of 6.06 ± 3.47. XGBoost achieved the best predictive performance (test R² = 0.451). SHAP analysis identified IRLS severity as the strongest predictor of PSQI across all three models, followed by EDSS score and depression in the Random Forest and XGBoost models, while daytime sleepiness (ESS) ranked consistently among the top predictors. Cardiometabolic and inflammatory parameters contributed inconsistently, with several showing effects opposite to physiological expectation. Conclusions: Sleep impairment in MS was driven mainly by restless legs syndrome severity and neurological disability, with depression also contributing in the two best-performing models. Disease-related and symptomatic factors outweighed cardiometabolic and inflammatory contributions. These findings support the value of interpretable machine learning for identifying the multifactorial drivers of sleep quality in MS.