1189 Machine Learning-Based Clinical Decision Support System (CDSS) for Efficient Sleep Clinic Management: Targeting OSA and RBD

Abstract Introduction The diagnosis of growing public health challenges, Obstructive Sleep Apnea (OSA) and REM Sleep Behavior Disorder (RBD), is hindered by Polysomnography's (PSG) inefficiencies: low-specificity screening and a laborious, subjective manual scoring process complicated by differentiating isolated RBD (iRBD) from OSA-induced "pseudo-RBD." This project proposes a two-stage, integrated Machine Learning (ML)-leveraged CDSS to overcome this diagnostic bottleneck and improve clinical management. Methods Retrospective data was collected from 1,256 patient records at Ewha University Medical Center Mokdong Hospital (April 2016 - March 2022) under IRB approval. Excluding patients with CPAP titration, split-night studies, age 40, or unclear diagnoses, the final cohort was 749 patients (62 RBD, 324 moderate-to-severe OSA, 15 coexisting, 348 other sleep disorders). The comprehensive dataset included demographics, lifestyle, 198 standardized questionnaire responses, and 49 manual PSG scoring variables. From 225 initial demographic and questionnaire items, 88 features were selected for machine learning classifier development. Results To address data deficiencies and optimize model efficacy, the dataset was partitioned into training (70%) and test (30%) subsets. Missing values within the training set were imputed utilizing the HyperImputer approach. This was succeeded by synthetic data augmentation employing TabSyn, which generated an additional 2,000 samples with a Cumulative Distribution Difference Error (CDDE) and Pairwise Correlation Consistency (PCC) exceeding 90%. A binary classification model was subsequently developed using the XGBoost gradient boosting algorithm. Model performance was validated through 5-fold cross-validation, achieving a sensitivity of 89% and a precision of 76%. Feature importance in the best-performing model was assessed using SHAP, permutation importance, XGBoost gain importance, and mutual information. Models constructed using the top 10, 20, and 30 features maintained performance comparable to the full-feature model (88 features), demonstrating a sensitivity of approximately 86% and a precision of 76%. Conclusion This CDSS is engineered to furnish objective metrics and standardize analysis, thereby mitigating inter-scorer variability and enhancing diagnostic precision across three critical phenotypes: pure iRBD, comorbid iRBD and OSA, and OSA presenting with parasomnia-like features. The anticipated outcome includes substantial clinical and economic benefits, notably a significant reduction in diagnostic expenditures. Support (if any) This study was supported by grants from the NRF funded by MSIT (RS-2024-00359247 to HJ Kim),