Scholarlink

Objective <p>Early and accurate prediction of neurological outcomes and mortality in comatose patients after cardiac arrest remains challenging. Multimodal data integrating heart and brain electrophysiological signals may improve prognostic accuracy, but distinct predictive patterns underlying neurological recovery versus survival are not well characterized.</p> Methods <p>We analyzed 331 patients from the International Cardiac Arrest Research (I-CARE) dataset. Synchronous electrocardiogram (ECG) and electroencephalogram (EEG) data [obtained within 72&#xa0;h after return of spontaneous circulation (ROSC)] and clinical variables were collected. Multimodal features were extracted from heart rate variability (time/frequency/nonlinear) and EEG (time/frequency/nonlinear/network topology), alongside heartbeat-evoked potentials (HEP). Features were temporally weighted by the time elapsed post-ROSC. Four machine learning models [logistic regression, support vector machine (SVM), random forest, and XGBoost] predicted neurological outcome [good (Cerebral Performance Category, CPC 1–2] vs. poor (CPC 3–5)] and mortality [survival (CPC 1–4) vs. death (CPC 5)]. Model development used tenfold cross-validation on the training set (80%), with hyperparameters optimized via Optuna, and performance was evaluated on a held-out test set (20%). Class imbalance was addressed via random oversampling and balanced class weights. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC–ROC), accuracy, sensitivity, specificity, precision, and the F1-score. SHapley Additive exPlanations (SHAP) analysis interpreted the optimal model and quantified feature contributions.</p> Results <p>XGBoost achieved the highest performance (AUC 0.95 for neurological outcome, 0.89 for mortality). SHAP analysis identified ShockableRhythm as the top predictor for neurological outcome and EEG kurtosis for mortality. Neurological outcome was more associated with EEG complexity and synchrony, whereas mortality was linked to ECG nonlinear dynamics and heart–brain coupling.</p> Conclusions <p>An XGBoost model integrating multimodal heart–brain electrophysiological features enables accurate early prognosis. The results reveal task-specific predictive patterns where neurological recovery shows stronger brain-centric associations, while survival shows stronger cardiac and heart–brain interaction associations. These findings provide a new direction for the development of personalized prognostic assessment tools based on multimodal physiological data.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Prediction of Neurological Outcome and Mortality in Cardiac Arrest Patients: An Explainable Machine Learning Study Integrating HRV, EEG, and Clinical Features

  • 期刊论文

  • 发表日期：

    2026年5月26日

查看全文
  • Shijing Wu,
  • Guojing Wang,
  • Ping Zhan,
  • Xiaohua Yu,
  • Weidong Wang,
  • Hongyun Liu

摘要

Objective

Early and accurate prediction of neurological outcomes and mortality in comatose patients after cardiac arrest remains challenging. Multimodal data integrating heart and brain electrophysiological signals may improve prognostic accuracy, but distinct predictive patterns underlying neurological recovery versus survival are not well characterized.

Methods

We analyzed 331 patients from the International Cardiac Arrest Research (I-CARE) dataset. Synchronous electrocardiogram (ECG) and electroencephalogram (EEG) data [obtained within 72 h after return of spontaneous circulation (ROSC)] and clinical variables were collected. Multimodal features were extracted from heart rate variability (time/frequency/nonlinear) and EEG (time/frequency/nonlinear/network topology), alongside heartbeat-evoked potentials (HEP). Features were temporally weighted by the time elapsed post-ROSC. Four machine learning models [logistic regression, support vector machine (SVM), random forest, and XGBoost] predicted neurological outcome [good (Cerebral Performance Category, CPC 1–2] vs. poor (CPC 3–5)] and mortality [survival (CPC 1–4) vs. death (CPC 5)]. Model development used tenfold cross-validation on the training set (80%), with hyperparameters optimized via Optuna, and performance was evaluated on a held-out test set (20%). Class imbalance was addressed via random oversampling and balanced class weights. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC–ROC), accuracy, sensitivity, specificity, precision, and the F1-score. SHapley Additive exPlanations (SHAP) analysis interpreted the optimal model and quantified feature contributions.

Results

XGBoost achieved the highest performance (AUC 0.95 for neurological outcome, 0.89 for mortality). SHAP analysis identified ShockableRhythm as the top predictor for neurological outcome and EEG kurtosis for mortality. Neurological outcome was more associated with EEG complexity and synchrony, whereas mortality was linked to ECG nonlinear dynamics and heart–brain coupling.

Conclusions

An XGBoost model integrating multimodal heart–brain electrophysiological features enables accurate early prognosis. The results reveal task-specific predictive patterns where neurological recovery shows stronger brain-centric associations, while survival shows stronger cardiac and heart–brain interaction associations. These findings provide a new direction for the development of personalized prognostic assessment tools based on multimodal physiological data.