Improved AI-based Prediction of ICU Mortality of Sepsis Patients based on Change in HRV Parameter upon Fluid Bolus Therapy
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Purpose
To examine the role of heart rate variability (HRV) changes in response to fluid bolus therapy (FBT) on mortality prediction among sepsis patients in the intensive care unit (ICU).
Methods
Adult sepsis patients in the Medical Information Mart for Intensive Care (MIMIC)-III and MIMIC-IV clinical database, who also had continuous ECG in the corresponding waveform database, within 48 hours of ICU admission were selected for the study. All models were developed using a common architectural template, each based on an ensemble of decision trees making use of eXtreme Gradient Boosting (XGBoost). A mortality prediction model (Model-1) was built including SAPS II score and common pool of features (CPOF) comprising of age, sex, ethnicity, insurance, admission type, heart rate, mean arterial pressure (MAP) and Elixhauser comorbidity score. A combination of the standard deviation of NN intervals (SDNN) as the HRV parameter, and FBT response in terms of the change in MAP ( Δ MAP) as well as the change in SDNN ( Δ SDNN) were then added hierarchically and also separately to Model-1 to develop additional models (Model-2 to Model-6). Finally, the performance of all the developed models were compared.
Results
A total of 5960 patients were screened and 542 were included following exclusion criteria. The model with a baseline traditional feature set (SAPS II, CPOF and Δ MAP) was surpassed by an augmented model (Model-6) using HRV measurements (SDNN and Δ SDNN) as additional features in terms of predictive performance (AUC: 0.871 vs 0.921, maximum F 1 score: 0.731 vs 0.792, indicating respective gains of 5.74% and 8.34%). In order of feature importance, following SAPS II score, Δ SDNN superseded Δ MAP.
Conclusions
The Δ SDNN can be an important predictor of mortality in septic patient requiring fluid bolus and is stronger than Δ MAP. Clinicians can choose Δ SDNN as an additional bedside parameter to predict mortality risk in sepsis patients.
