Machine Learning Based Predictions of Adverse Events in Prone Positioned Patients with Acute Respiratory Distress Syndrome (ARDS)

ABSTRACT Introduction. Acute Respiratory Distress Syndrome (ARDS) represents one of the primary causes of acute respiratory failure in intensive care, often requiring advanced respiratory support strategies. Prone positioning of the ventilated patient is a widely validated maneuver; however, it is burdened by a non-negligible risk of complications. In this context, Artificial Intelligence (AI) is proposed as a decision-support tool capable of integrating a large amount of clinical data to optimize the management of critical patients. Materials and Methods. This multicenter observational study utilized data from the international PROVENT-C19 registry, including adult patients with confirmed COVID-19 undergoing invasive mechanical ventilation and at least one prone positioning session in the ICU. Data were collected in a standardized manner via the REDCap platform. The primary endpoint was the occurrence of any complication in the ICU, while secondary endpoints included life-threatening complications and pressure sores. Dedicated predictive models were developed for each outcome using an ensemble machine learning approach based on Super Learner, combining parametric and non-parametric algorithms. Missing data were managed through multiple imputation with chained equations. Model performance was evaluated via internal validation with 10-fold cross-validation, primarily estimating discriminative capacity through the cross-validated area under the ROC curve and also analyzing the model's calibration. Results. Among the 1,722 patients included in the study, 68.9% presented at least one complication during their ICU stay and 59.2% developed life-threatening complications; pressure sores occurred in about a quarter of the cases. In the 10-fold cross-validation, the Super Learner showed modest discriminative capacity for all outcomes, with a cross-validated AUROC of 0.67 (95% CI: 0.64–0.70) for any complication, 0.64 (95% CI: 0.61–0.67) for life-threatening complications, and 0.61 (95% CI: 0.58–0.64) for pressure sores. The performance of the Super Learner ensemble, the Discrete Super Learner, and 11 candidate algorithms were compared using cross-validated mean squared error (cv-MSE) and cross-validated AUROCs (cv-AUROC). Overall differences between the models were limited. The Super Learner showed the best discriminative capacity for the "any complication" and "life-threatening complication" outcomes (cv-AUROC 0.67 and 0.64), with performance comparable to the best individual models. For pressure sores, discrimination was lower (cv-AUROC 0.61) without a clear advantage of the ensemble over the individual algorithms, while XGBoost showed the worst performance and the reference model presented random discrimination (cv-AUROC = 0.50). Conclusions. Predictive models based on Super Learner showed an overall modest discriminative capacity, insufficient for direct clinical use; however, the integration of dynamic information, healthcare process variables, and rigorous external validation could improve risk prediction and support targeted preventive strategies.