Frailty-Driven Prediction of Inpatient Sleep Disorder Diagnoses with Explainable AI

Sleep disorders affect a substantial proportion of hospitalized patients yet remain among the most systematically underdiagnosed conditions in acute care medicine, with up to 80% of moderate-to-severe cases carrying no formal diagnosis at the time of admission. At the same time, frailty—a state of heightened physiological vulnerability arising from cumulative multi-system biological decline—is present in 40–80% of inpatients and shares deep, bidirectional neurobiological pathways with sleep pathology through shared mechanisms of circadian dysregulation, hypothalamic-pituitary-adrenal axis activation, and chronic low-grade inflammation. Despite this convergence, no study has integrated validated, administratively computable frailty phenotyping with a machine learning framework specifically designed to predict inpatient sleep disorder diagnosis at the point of hospital admission. To address this gap, we developed and evaluated a suite of five binary classification models—XGBoost, Random Forest, LightGBM, CatBoost, and Decision Tree—using 9,682 balanced hospitalization episodes from the MIMIC-IV (version 2.2) database. The predictor set comprised 23 admission-time structured features across three domains: frailty and comorbidity burden, including the Hospital Frailty Risk Score (HFRS) derived from ICD-10 codes, the Elixhauser comorbidity index, prior admission history, and six binary disease flags (obesity, hypertension, type 2 diabetes, heart failure, COPD, and depression/anxiety); physiological and laboratory biomarkers from the first 24 hours of care, including minimum SpO₂, heart rate variability, hemoglobin, creatinine, albumin, and arterial blood gas parameters; and sociodemographic and administrative variables encompassing age, sex, ethnicity, insurance type, and admission acuity. Two binary outcomes were modeled independently: any sleep disorder diagnosis (ICD-10: G47.x) and insomnia specifically (ICD-10: G47.00). Model performance was assessed through five-fold stratified cross-validation and bootstrap confidence intervals (n = 1,000 iterations), with predictor importance quantified using SHapley Additive exPlanations (SHAP). XGBoost achieved the strongest aggregate performance across all evaluation metrics, attaining an area under the receiver operating characteristic curve (AUC) of 0.871 (95% CI: 0.856–0.887), accuracy of 79.6%, F1-score of 0.820, and sensitivity of 94.9%, correctly identifying 903 of 952 true positive cases in the held-out test set; all gradient boosting frameworks substantially outperformed the Decision Tree baseline (AUC 0.836). SHAP analysis identified the HFRS and Elixhauser index as the two dominant predictors, followed by depression/anxiety, obesity, hypertension, and minimum SpO₂—a pattern that is mechanistically consistent with established pathophysiological literature on frailty-associated sleep pathology. The well-calibrated probability outputs of the XGBoost model make it directly suitable for integration into clinical decision support systems, offering a deployable, interpretable screening tool for inpatient sleep disorder identification that requires no dedicated instrumentation beyond routine admission data.