0367 BREATH-Biomarkers of Respiratory Exacerbation in Asthmatic Children Through Sleep Health

Abstract Introduction Asthma is a chronic lung disease affecting 6.5–10% of children and severe asthma exacerbations are often preceded by a period of poor asthma control, suggesting a window for preventive intervention. There is no reliable method to predict asthma exacerbation. Sleep disruption is a common complaint among children with asthma. Polysomnography (PSG) is routinely used in this population to assess for sleep disordered breathing and captures brain, heart, and respiratory signals for 6-9 hours. Our preliminary data suggests that sleep fragmentation, noted on PSG, is associated with severe asthma exacerbation within the following year. However, till date, sleep microstructures have not been studied to identify the children at risk for asthma exacerbation. Methods Retrospective cohort analysis to identify correlates and quantify the prognostic utility of poor sleep microstructures, utilizing machine learning approaches, associated with increased risk of severe asthma exacerbation event (SE) 12-months post a sleep study. We extracted microstructural features, including spectral band power statistics, interhemispheric coherence, and temporal statistics, from multi-modality physiological signals across sleep stages (wake, N1, N2, N3, REM). We implemented two machine learning models (e.g., XGboost and Lasso) with 10-fold cross-validation. We applied Explainable AI (i.e., SHAP) to identify the most predictive microstructural features for SE risk. Results The study cohort included 160 patients (36% female, 33% African American, mean age of 10 ± 4 years). The XGBoost model demonstrated good discriminative performance for predicting 1-year SE risk (C statistic=0.72 ± 0.14). SHAP analysis revealed that gamma-band interhemispheric EEG coherence (T5–T6) during wake was the strongest predictor (mean |SHAP| = 0.57 ± 0.15). Other contributory features included delta-band EEG kurtosis in N1 (0.38 ± 0.11), a temperature-derived flow limitation index during wake (0.29 ± 0.18), and beta-band EEG kurtosis in N3 (0.29 ± 0.14). Electromyogram (EMG) and Eelectrooculogram (EOG) features also contributed substantially, e.g., including N1 high-frequency EMG power and N1 REM-related EOG power. Conclusion Sleep microstructural features that typically cannot be scored manually are different in children with and without asthma exacerbation within the following year. These microstructure measures may signify subclinical features of sleep fragmentation and respiratory effort disruption and offer accurate prognosis for targeted prevention interventions. Support (if any)