Real time automatic risk prediction in ICU patients treated with ECMO

The COVID-19 pandemic has underscored the relevance of intensive care medicine. ICUs faced repeated strain, relying on advanced equipment for care provision. ECMO emerged as pivotal, providing support to patients lacking sufficient oxygen delivery via invasive mechanical ventilation. This device consistently generates high flows of multidimensional data, showcasing the potential of leveraging machine learning-based systems. Considering 81 labeled multivariate time series referring to patients with COVID-19 pneumonia treated with ECMO support, a Support Vector Machine with varying kernel functions and a Random Forest model were employed to discern clinical condition deterioration and improvement. Performance assessment resulted in the selection of the optimal Random Forest model, presenting exceptional metrics and precise calibration. The Clinical Assessment Score, computed based on the selected model, provided a real-time assessment of the patient's clinical dynamics, effectively anticipating significant clinical condition deterioration and improvement, achieving an AUROC of 0.9176, 0.8944, and 0.8556 considering preceding time intervals of 4, 8, and 12 hours, respectively.