Frailty-Driven Prediction of Inpatient Obstructive Sleep Apnea and Related Sleep Disorder Diagnoses Using Explainable...

Area of Science:

• Computational medicine and health informatics
• Sleep medicine and respiratory disorders
• Geriatrics and internal medicine

Background:

• Obstructive sleep apnea (OSA) and frailty are highly prevalent in hospitalized patients, sharing common pathophysiological pathways.
• A significant diagnostic gap exists for OSA and other sleep disorders among inpatients, leading to delayed treatment and adverse outcomes.
• Current clinical practice lacks tools to predict inpatient sleep disorders at hospital admission using readily available data.

Purpose of the Study:

• To develop and validate an explainable machine learning framework for predicting inpatient sleep disorder diagnoses, focusing on OSA, at the time of hospital admission.
• To integrate administratively computable frailty phenotyping with clinical and demographic data for predictive modeling.
• To create a tool that can be integrated into clinical decision support systems for early identification of at-risk patients.

Main Methods:

• Utilized the MIMIC-IV database (v2.2) with 9682 hospitalization episodes for model development and evaluation.
• Developed five binary classification models (XGBoost, Random Forest, LightGBM, CatBoost, Decision Tree) using 23 admission-time features including frailty scores (HFRS, Elixhauser), comorbidities, physiological data (SpO2), and demographics.
• Assessed model performance using five-fold cross-validation and quantified predictor importance with SHapley Additive exPlanations (SHAP).

Main Results:

• The XGBoost model demonstrated superior performance with an AUC of 0.871, accuracy of 79.6%, and sensitivity of 94.9%.
• SHAP analysis revealed Hospital Frailty Risk Score (HFRS) and Elixhauser index as key predictors, indicating the model captures the OSA-frailty axis.
• The model's predicted probabilities were well-calibrated across all risk deciles, suggesting reliable risk estimation.

Conclusions:

• Admission-time frailty and clinical data can effectively predict inpatient sleep disorders, primarily OSA, with high accuracy and reliability.
• The interpretable XGBoost model is suitable for clinical decision support, requiring only routine admission data for screening.
• This framework facilitates early diagnosis and treatment initiation, potentially reducing diagnostic gaps, perioperative risks, and adverse outcomes in frail hospitalized patients.