An Ensemble Learning Approach for Predicting Acute Respiratory Distress Syndrome Within Polytraumatic Patients In Imbalanced Dataset

Acute Respiratory Distress Syndrome (ARDS) is a severe, often overlooked complication in polytrauma patients, complicated by its varied symptoms and delayed diagnosis through imaging. To address this, our study developed a machine learning tool to predict ARDS early, using data from 407 polytrauma patients collected at admission, 12 hours, and 24 hours later. We aimed to create a clear, practical decision-support tool for clinicians. For this, we retrospectively analyzed de-identified patient data from a clinical observational group. To improve predictive accuracy, we handled data imbalance and selected the most informative clinical features. A Random Forest classifier was trained and evaluated with metrics like AUROC, PR-AUC, F1 score, recall, and precision, while SHAP values helped explain the model’s predictions. The model performed strongly, achieving an AUROC of 0.96, PR-AUC of 0.81, recall of 0.92, and F1 score of 0.76. Key predictors included arterial oxygen pressure (PaO2), flail chest injury, blood pH at 24 hours, hemothorax, and total fluid resuscitation, offering clinicians actionable insights for early ARDS detection. This ML model permits early prediction of ARDS in polytraumatic patients, providing potential for eventual incorporation into electronic health records to minimize mortality through in-time interventions.