Predicting the clinical evolution of septic patients from routinely collected data and vital signs variability using machine learning.

Objective: The existing literature lacks a comprehensive analysis of the clinical evolution of septic patients, which is highly heterogeneous and patient-dependent. The aim of this study is to develop machine learning models capable of predicting the clinical evolution of septic patients and to evaluate the predictive ability of features.

Approach: Data from intensive care unit (ICU) septic patients were extracted from the freely available HiRID database and a comprehensive pipeline for time series analysis of critical care data was developed. Predictive models of cardiovascular deterioration (based on mean pressure and lactate values) and global organ dysfunction (based on SOFA score) were developed, and the addition of variability, such as entropies, cross-entropies and cross-correlation of heart rate and blood pressure, was tested against the use of standard metrics alone.

Main results: The best model achieved an area under the ROC curve value of 0.9671, with SOFA score values and trends being the most important features in the model, followed by features related to lactate, fluid balance, therapy and entropy values of blood pressure.

Significance: The results show that the dynamics of vital signs and their cross-coupling, as captured by the proposed variability indices, can provide additional insights into the physiological responses to the therapy administered.