Go to IEEE EMBS BHI 2025 Conference homepageA Deep Counterfactual Framework for High-Flow Nasal Cannula and Non-Invasive Ventilation Recommendations for Acute Respiratory Failure
Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.
Keywords: Counterfactual Inference, Individualized Treatment Effect, Respiratory Failure, High-Flow Nasal Cannula, Non-Invasive Ventilation
TL;DR: RepFlow-CFR is a deep counterfactual model that personalizes the choice between HFNC and NIV in acute respiratory failure.
Abstract: High-flow nasal cannula (HFNC) and non-invasive ventilation (NIV) are commonly used respiratory support therapies for acute respiratory failure (ARF). However, current randomized trials provide limited guidance for individualized treatment decisions in this patient population. We propose RepFlow-CFR, a deep counterfactual inference model designed to estimate the individualized treatment effects (ITEs) of HFNC versus NIV. The model was applied to retrospective data from ICU cohorts at two independent health systems, UC San Diego (UCSD) Health and UC Irvine (UCI) Health. The primary outcome was the need for invasive mechanical ventilation (IMV). After adjusting for confounders, a multivariable logistic regression analysis at the UCSD site showed that concordance with the RepFlow-CFR model's recommendations was significantly associated with a lower risk of IMV. Specifically, the odds ratio (OR) for IMV was 0.661 (p$<$0.001) for concordance with a NIV recommendation and 0.677 (p$=$0.019) for concordance with an HFNC recommendation. These results demonstrated a more significant and consistent protective effect compared to baseline methods like Causal Forest and X-learner. The findings underscore the model's potential to provide data-driven, personalized guidance for respiratory support decisions in critically ill patients.
Track: 4. Clinical Informatics
Registration Id: D7NFKTXYY7Q
Submission Number: 232
Loading