Predicting Adverse Events in the Cardiothoracic Surgery Intensive Care Unit Using Machine Learning: Results and Challenges

Abstract

Abstract It is highly important to anticipate impending problems in patients in the cardiothoracic intensive care unit (CTICU) and be proactive with respect to prediction of adverse events, enabling interventions to prevent them. In order to develop models that predict the occurrence of adverse events after cardiac surgery, a dataset of 9,237 patients was constructed of a single center’s Society of Thoracic Surgeons (STS) internal database. 1,383 of those patients had developed at least one of seven defined adverse events for this analysis. For the control set, we randomly picked 1,383 patients from the group who did not develop any adverse event. The ensemble learning algorithm, random forest, was applied and outperformed the best reported logistic regression models for similar task (c-statistic of ∼0.81), by achieving an AUC of 0.86 with a 95% CI of [0.81-0.90], specificity of 0.72, sensitivity of 0.82, PPV of 0.78 and NPV of 0.77. In the future, we plan to run a similar evaluation process on a multicenter dataset, and then use this static prediction model as a context for using time-evolving data to develop algorithms for real-time feedback to care teams. In acute care settings, such as the operating room and intensive care unit, the ability to anticipate potentially fatal complications will be enhanced by using supervised machine learning algorithms.

Ähnliche Arbeiten