A55-41 Temporal Explainability in Discrete-time Survival Modeling for Acute Respiratory Distress Syndrome (ARDS) on Mechanical Ventilation: Comparing Clinical Trends and Chest X-ray Embeddings

Abstract Rationale Patients with ARDS requiring mechanical ventilation have 30-50% mortality, yet trajectories evolve dynamically. Traditional survival models rely on baseline features and do not incorporate evolving physiology or high-dimensional imaging. We developed a machine learning discrete-time survival model to a) integrate multimodal features including chest X-rays, b) incorporate temporal clinical trends, and c) assess time-varying feature importance. Methods We used the MIMIC-IV and MIMIC-CXR-JPG datasets to identify mechanically ventilated patients with ARDS using ICD diagnosis codes. Inclusion criteria were: age18, duration of mechanical ventilation 48 hours, and at least one chest X-ray obtained during mechanical ventilation. For all laboratory and vital sign variables, we engineered temporal features including the mean, standard deviation, and slope of rolling windows (3-day and 7-day), as well as cumulative mean and slope since ventilation initiation. After splitting the dataset into training/testing sets (80:20) by patients, chest X-ray embeddings were extracted with DenseNet121, and reduced to 166 dimensions using principal component analysis (PCA). An Extreme Gradient Boosting (XGBoost) discrete-time survival model was trained with a 10% internal validation subset for hyperparameter tuning (e.g., learning rate, estimators, tree complexity, and regularization). Two models were developed- with and without X-ray embeddings- to evaluate incremental benefit of radiography. Model explainability was assessed using SHAP (Shapley Additive Explanations) values. Results A total of 1,319 patients with 10,511 chest X-rays were included. Model performances for features with and without X-ray embeddings were comparable (AUROC 0.93, AUPRC 0.43 vs. AUROC 0.93, AUPRC 0.42, respectively). Survival curves with Kaplan-Meier estimates showed that both models overestimated survival probabilities in later ICU days. SHAP analysis revealed that vital and ventilator features consistently dominated feature importance, with X-ray embeddings contributing the least over time. Conclusions We developed a machine learning-based discrete-time survival model to predict daily mortality risk among ARDS patients receiving mechanical ventilation. Although chest X-ray embeddings provided additional prognostic information, most predictive value was derived from physiologic and laboratory trends. Performance declined slightly late in the ICU course, suggesting opportunities for improved modelling of prolonged critical illness. This abstract is funded by: None