A Machine Learning Algorithm to Predict Hypoxic Respiratory Failure and risk of Acute Respiratory Distress Syndrome (ARDS) by Utilizing Features Derived from Electrocardiogram (ECG) and Routinely Clinical Data

Abstract

The recognition of Acute Respiratory Distress Syndrome (ARDS) may be delayed or missed entirely among critically ill patients. This study focuses on the development of a predictive algorithm for Hypoxic Respiratory Failure and associated risk of ARDS by utilizing routinely collected bedside monitoring. Specifically, the algorithm aims to predict onset over time. Uniquely, and favorable to robustness, the algorithm utilizes routinely collected, non-invasive cardiorespiratory waveform signals. This is a retrospective, Institutional-Review-Board-approved study of 2,078 patients at a tertiary hospital system. A modified Berlin criteria was used to identify 128 of the patients to have the condition during their encounter. A prediction horizon of 6 to 36 hours was defined for model training and evaluation. Xtreme Gradient Boosting algorithm was evaluated against signal processing and statistical features derived from the waveform and clinical data. Waveform-derived cardiorespiratory features, namely measures relating to variability and multi-scale entropy were robust and reliable features that predicted onset up to 36 hours before the clinical definition is met. The inclusion of structured data from the medical record, namely oxygenation patterns, complete blood counts, and basic metabolics further improved model performance. The combined model with 6-hour prediction horizon achieved an area under the receiver operating characteristic of 0.79 as opposed to the first 24-hour Lung Injury Prediction Score of 0.72.