Clinical Features and Physiological Signals Fusion Network for Mechanical Circulatory Support Need Prediction in Pediatric Cardiac Intensive Care Unit.

Abstract

We link the hemodynamic response to inotropic agents with outcomes related to Mechanical Circulatory Support (MCS) by analyzing physiological time series and clinical features using a Machine Learning/Deep Learning ensemble approach for multi-modal waveforms in the pediatric cardiac intensive care setting of a quaternary-care hospital. Unlike existing studies that typically process a single feature type or focus on short-term diagnoses from physiological signals, our novel system processes minute-by-minute multi-sensor data to identify the need for MCS in patients admitted with acute decompensated heart failure. The data used includes tabular clinical features, time series from hemodynamic monitors, and raw waveforms from electrocardiogram and arterial blood pressure signals. Our predictions support an early identification of high-risk patients after just two days of Intensive Care Unit (ICU) admission, with classification and feature importance results confirming the predictive ability of the early hemodynamic response to inotropic agent administration, achieving an AUC of 0.88 in the prediction classification task. This is particularly significant in cases where clinical decisions are not straightforward, such as those in the cohort for this study.