Abstract
Aim: This study aimed to develop and evaluate a machine-learning model capable of predicting cardiogenic shock progression using continuously recorded hemodynamic telemetry data. Methods: Patients with cardiogenic shock undergoing continuous hemodynamic monitoring in a cardiac intensive care unit were included. A gradient-boosting model used features representing the level, trend, and variability of monitored parameters to predict progression to refractory shock, mechanical circulatory support, or death. Results: Of 1,600 treated patients, 33% went on to progress. The machine-learning model had an area under the curve value of 0.86 (95% CI 0.82 to 0.89), and the analysis indicated 85% sensitivity and 90% negative predictive value, with good performance of the model. Trends of lactate, vasoactive-inotropic score and cardiac power output proved to be the most influential predictor variables and above any single static variable; furthermore, performance was consistent across shock stage, aetiology and age. Conclusion: Continuous monitoring of trouble in a machine-learning model can predict cardiogenic shock progression with great accuracy; the anti-temporal trends in the continuous monitoring data were the most predictive of cardiogenic shock progression, highlighting the importance of continuous monitoring in the cardiac intensive care unit for automated early warning.
Keywords: Cardiogenic shock, Machine learning, Hemodynamic monitoring, Cardiac intensive care