BASH-GN

BASH-GN classifies individuals into obstructive sleep apnea (OSA) risk categories using machine-learning-derived questionnaire models that incorporate OSA subtypes to improve screening accuracy.

Key Features:

• Development cohorts: Model development used data from the Sleep Heart Health Study Visit 1 (SHHS 1) and the Wisconsin Sleep Cohort (WSC).
• Risk threshold: High-risk classification is defined using an apnea-hypopnea index (AHI) threshold of ≥15 events per hour.
• Feature ranking: Mutual information was used to rank potential risk factors and the top 50% of features were selected for analysis.
• Phenotype stratification: Participants are classified into low and high phenotype groups based on calculated risk scores.
• Classifier architecture: Two logistic regression classifiers were trained, each tailored to predict different OSA risk subtypes.
• Benchmarking: Performance was compared against the Four-Variable, Epworth Sleepiness Scale, Berlin, and STOP-BANG questionnaires.
• Performance metrics: Evaluated on independent test sets (SHHS 1 n=1237, WSC n=1120) with AUROC of 0.78 (SHHS 1) and 0.76 (WSC) and AUPRC of 0.72 (SHHS 1) and 0.74 (WSC).

Scientific Applications:

• Screening and triage: Prioritizes identification of individuals at high risk for OSA for further diagnostic testing or intervention based on AHI-defined risk.
• Subtype-aware risk assessment: Incorporates OSA subtypes into risk estimation to address heterogeneity in OSA presentations.
• Comparative evaluation: Serves as a machine-learning benchmark against established screening questionnaires in sleep medicine research.

Methodology:

Model development used SHHS 1 and WSC cohort data; mutual information ranked risk factors with the top 50% selected; participants were stratified into low/high phenotype groups by risk score; two logistic regression classifiers were trained and evaluated on independent test sets using AUROC and AUPRC and compared to Four-Variable, Epworth Sleepiness Scale, Berlin, and STOP-BANG, with high-risk defined as AHI ≥15 events/hour.