Real time automatic risk prediction in ICU patients treated with ECMO

Filipe Miguel de Oliveira Ribeiro

Key information

Authors:

Filipe Miguel de Oliveira Ribeiro (Filipe Miguel de Oliveira Ribeiro)

Supervisors:

João Miguel Raposo Sanches (João Miguel Raposo Sanches); Ana Catarina Fidalgo Barata (Ana Catarina Fidalgo Barata)

Published in

11/27/2024

Abstract

Extracorporeal Membrane Oxygenation (ECMO) is a therapeutic intervention employed in intensive care medicine that provides life support to critically ill patients whose lungs and heart function are severely compromised and proved decisive during the COVID-19 pandemic. ECMO support relies on a complex network of technologically advanced systems that monitor the patient’s clinical condition throughout hospitalization, generating multidimensional and multidomain datasets. Assessing ECMO datasets is challenging, presenting a promising opportunity for applying Machine Learning (ML) techniques. Using 81 labeled Multivariate Time Series (MTSs) from patients with COVID-19 pneumonia treated with ECMO support, a Support Vector Machine (SVM) with varying kernel functions and a Random Forest model was trained to distinguish between clinical deterioration and improvement throughout hospitalization. The Random Forest model achieved the best predictive performance and calibration. Ultimately, feature importance analysis provided insights into its predictions, enhancing interpretability and practical applicability. The Random Forest model with 100 trees was then used to compute a risk score, which provided a real-time estimate of the risk of clinical deterioration throughout hospitalization under ECMO support for each patient. The score effectively anticipated periods of clinical decline and improvement, achieving an Area Under the Receiver Operating Characteristics (AUROC) curve of 0.9176, 0.8944, and 0.8556 for time intervals preceding these periods of 4, 8, and 12 hours, respectively. This study demonstrated the pivotal role of ML systems in supporting physicians to assess complex patient cohorts, uncovering insights that might have otherwise remained undetectable.