Disclosure(s): No relevant financial relationship(s) to disclose.
First Author: Hanna Hieromnimon, N/A Co-Author: Kimberly Coston, BSN, RN – Registered Nurse, Duke University Hospital Co-Author: Jared Houghtaling, N/A – Assistant Professor, Tufts Medical Center Co-Author: Yuriy Bronshteyn, MD, FASE – Associate Professor, Duke University School of Medicine, Duke University Health System
Introduction: Post-operative mechanical ventilation (POMV) has significant impacts on outcomes, workflow, and resource allocation. Current existing prediction models often utilize disparate variables, traditional regression techniques, single datasets, and rarely explore optimal feature subsets. We aim to identify key pre-operative predictors of POMV using interpretable machine learning (ML) methods on large, diverse datasets.
Methods: We conducted a retrospective cohort study using four critical care databases (MIMIC-IV (n=20122/8631) , eICU (n=7124/3055), CHoRUS (n=255/104), and INSPIRE (n=3955/1692)) with a 70/30 train/test split. Adult surgical patients were included with a primary outcome of need for POMV (binary). POMV was defined as the need for mechanical ventilation 24 hours post surgery. We evaluated features identified in other POMV models in addition to pre-op serum bicarbonate utilizing multiple ML modes (e.g. Random Forest, Decision Tree, Logistic Regression, FasterRisk) to predict POMV. Feature importance and model reduction techniques were used to identify parsimonious models.
Results: ML models demonstrated reasonable performance for predicting POMV across different datasets. Preliminary analysis showed comparable performance between tested ML algorithms (AUC0.731-0.587). Pre-op serum bicarbonate emerged as a potentially novel predictor with a possible non-linear (U-shaped) relationship with POMV risk. Feature importance analysis revealed a reduced set of features could retain substantial predictive power. A Random Forest model utilizing the top 4 pre-op features achieved an AUC of 0.688 compared to 0.690 for our full feature set. Interpretability techniques such as decision trees provided insight into model decision-making pathways.
Conclusions: Using interpretable ML models, pre-op data can predict POMV needs across a diverse population. Pre-op serum bicarbonate may be an under-recognized predictor warranting further investigation. Importantly, a parsimonious model with ≤5 features retains significant predictive capability, suggesting potential for a development of a simplified bedside tool for risk stratification and resource planning. Further work is needed to refine key features and assess the modifiability of identified risk factors.
