Enhancing Hypoglossal Nerve Stimulation Outcomes in Obstructive Sleep Apnea: A Machine Learning Candidate Selection Tool

Introduction: Obstructive sleep apnea (OSA) significantly impacts public health, yet adherence to continuous positive airway pressure (CPAP) therapy remains low. This study aims to develop and validate a machine learning–based clinical decision support tool (CDST) to optimize candidate selection for hypoglossal nerve stimulation.

Methods: This is a retrospective study analyzing data from 2013–2023 on 842 adults with moderate-to-severe OSA (AHI 15–65) treated with hypoglossal nerve stimulation (HNS). Data were sourced from peer-reviewed publications via PubMed, EMBASE, and Cochrane Library, with synthetic datasets generated through Monte Carlo simulations to address missing individual-level data (35% synthetic). Independent variables included demographics (age, gender, BMI), polysomnographic metrics (baseline AHI, oxygen desaturation index [ODI], Epworth Sleepiness Scale [ESS]), and anatomical factors from drug-induced sleep endoscopy (concentric velum collapse, lateral wall collapse). Preliminary analyses included multiple imputation, outlier detection, and normalization. A random forest classifier was trained on 80% of the data, validated on 20%, and assessed for accuracy, sensitivity, specificity, and AUC-ROC. The final model was deployed in a Python-based clinical decision support tool (CDST) for real-time HNS suitability predictions.

Results: The random forest model achieved 87% accuracy, 90% sensitivity, 84% specificity, and an AUC-ROC of 0.92. Significant predictors of HNS success were baseline AHI, BMI, concentric velum collapse, and ODI. Validation against published outcomes showed high consistency, with patients having high predicted success probabilities demonstrating substantial postoperative AHI improvements.

Conclusions: A Machine learning–based CDST offers clinicians a robust data driven tool to input patient-specific data and receive immediate, probability-based HNS success predictions, enhancing decision-making and optimizing patient outcomes.