Prediction of mortality in intensive care unit with short-term heart rate variability: machine learning-based analysis of the MIMIC-III database
Prognosis prediction in the intensive care unit (ICU) traditionally relied on physiological scoring systems based on clinical indicators at admission. Electrocardiogram (ECG) provides easily accessible information, with heart rate variability (HRV) derived from ECG showing prognostic value. However,...
Gespeichert in:
Veröffentlicht in: | Computers in biology and medicine 2025-03, Vol.186, p.109635, Article 109635 |
---|---|
Hauptverfasser: | , , , , , |
Format: | Artikel |
Sprache: | eng |
Schlagworte: | |
Online-Zugang: | Volltext |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Zusammenfassung: | Prognosis prediction in the intensive care unit (ICU) traditionally relied on physiological scoring systems based on clinical indicators at admission. Electrocardiogram (ECG) provides easily accessible information, with heart rate variability (HRV) derived from ECG showing prognostic value. However, few studies have conducted a comprehensive analysis of HRV-based prognostic model against established standards, which limits the application of HRV's prognostic value in clinical settings. This study aims to evaluate the utility of HRV in predicting mortality in the ICU. Additionally, we analyzed the applicability and interpretability of the HRV-integrated clinical model and identified the HRV factors that are most significant for patient prognosis.
A total of 2838 patients from the MIMIC-III database were retrospectively included in this study. These patients were randomly divided into training and testing sets at a 4:1 ratio. We collected 86 HRV indicators from patients’ lead II ECG readings between 0.5h-2h before the time of death in the ICU of deceased patients or time of discharge from the ICU of alive patients, in addition to 9 clinical parameters upon admission. Subsequently, machine learning models were developed by algorithms including logistic regression (LR), Random Forest (RF), Adaptive Boosting (Adaboost), Gradient Boost (GB), eXtreme Gradient Boosting (XGB), and Light GBM (LGB) algorithms. An ensemble model that integrated these six algorithms, along with a deep neural network model, was also explored. The ten most important variables were identified using the Shapley method. Subsequently, an HRV-modified clinical scoring system was constructed through recursive feature elimination.
The study demonstrated that the integrated model, utilizing both clinical and HRV features, outperformed the model based solely on clinical information in XGB, LGB and LR algorithms (p=0.005∼0.03). The ensemble model exhibited the best performance (AUROC=0.878), followed closely by XGB algorithm (AUROC=0.869). Both of these models significantly outperformed the APS III scoring system (AUROC=0.765). Notably, this improvement is not dependent on a specific disease but rather on the timing of ECG recordings that are closer to clinical endpoints. For parameter analysis, Shapley's method identified MSEn, SD1SD2, DFAα1, and DFAα2 as key HRV features in predicting mortality. These variables also showed significant differences in univariate analysis across patients with diffe |
---|---|
ISSN: | 0010-4825 |
DOI: | 10.1016/j.compbiomed.2024.109635 |