B110-07 Quantifying Breathing Pattern Disorders Using a Machine Learning Approach to Cardiopulmonary Exercise Testing: A Study of a National Cohort of Deployed Veterans

Abstract Rationale Dysfunctional breathing (DB) is a multifactorial condition characterized by erratic breathing patterns, particularly during exercise, causing breathlessness disproportionate to underlying pathology. While DB affects up to 9.5% of the population, no gold-standard diagnostic protocol exists. Recent studies have proposed using select cardiopulmonary exercise testing (CPET) parameters to identify DB. Current criteria for DB, on CPET, include the presence of 1 or more of the following: 1) inefficient ventilation (VE/VCO235), 2) depressed end-tidal CO2 (30mmHg), and 3) breathing pattern disorders (BPD) - ‘erratic’ tidal volume (VT) and/or respiratory frequency (Rf). The latter criteria lack objectivity and rely on expert raters to define its presence. This study evaluates a novel objective approach to defining BPD in a sample of military veterans with respiratory complaints who are at high risk of DB. Methods A quantile random forest model (QRF) was developed to predict expected VT and Rf during exercise. The QRF was trained using a reference cohort of 45 non-treatment-seeking Veterans (87% white, 89% male; meanSD: Age=448 years, Height=693 in, Weight=19328 lbs) with no visual signs of BPD. Two separate models were trained, one for Rf (root mean square error (RMSE)=4.04, model-fit R²=0.75) and one for VT (RMSE=0.39, R²=0.77). The 90% prediction intervals (PI) generated from the QRF was used to calculate two coverage rates (CR) (percentage of observed Rf and VT within the PI from submaximal to maximal exercise). A CR 70% for either parameter was selected to define BPD. CR calculations were applied to an independent cohort of 60 symptomatic Veterans (77% white, 87% male; meanSD: Age=469 years, Height=684 in, Weight=21346 lbs) with exertional dyspnea enabling objective assessment of BPD and identification of DB (Figure). Logistic regression adjusted for age, sex, height, and weight was used to examine associations between BPD and dyspnea during exercise. Results Thirty-five Veterans (58%) met criteria for DB (meanSD coverage%: VT = 69.827.1 L, Rf = 77.522.0 bpm), among which 15 (43%) had an isolated BPD. Logistic regression analysis comparing Veterans with isolated BPD to Veterans without DB showed a non-significant relationship between peak dyspnea and BPD (OR: 1.4395%CI: 0.99-2.22, p = 0.07). Conclusions Dysfunctional breathing was highly prevalent among Veterans evaluated for exertional dyspnea and was primarily driven by breathing pattern disorder. The quantile random forest model provided an objective and quantitative method for detecting abnormal tidal volume and breathing frequency variability. Future studies should validate this approach against clinician-rated assessments and determine its diagnostic and prognostic implications. This abstract is funded by: None