Prognostic value of heart rate variability for neurological outcomes in cardiac arrest patients: machine learning-based analysis of the International Cardiac Arrest Research (I-CARE) database.

Objective: To predict neurological outcome of cardiac arrest patients who are in a comatose state, incorporating Heart Rate Variability (HRV) from electrocardiogram (ECG) signals to enhance prognostic value. Methods: The International Cardiac Arrest Research (I-CARE) database which includes clinical information and continuous ECG data for 1020 patients aged 16 or older who were comatose after return of spontaneous circulation following cardiac arrest. We collected 86 HRV indicators from patients' ECG readings, in addition to clinical parameters, to predict neurological outcome [categorized into good (Cerebral Performance Category (CPC) score 1-2) and poor (CPC score 3-5)]. Six machine learning models were constructed to predict neurological outcome, including logistic regression (LR), Random Forest (RF), Adaptive Boosting (Adaboost), Gradient Boost (GB), eXtreme Gradient Boosting (XGB), and Light GBM (LGB). A rigorous 5 repetitions of 10-fold cross-validation was implemented as the internal validation strategy on the training set to mitigate overfitting, and an ensemble model integrating these six validated algorithms was further explored. Twenty most important variables were identified using the Shapley method. The correlation between neurological outcomes and different HRV metrics were analyzed. Recursive feature elimination was applied to develop a clinical scoring system modified by HRV indicators. Results: A total of 406 individuals with complete ECG & HRV data were included. We used 70% data as training set, with the rest 30% as test set. 5 × 10-fold cross-validation confirmed the robustness of all models, with the XGB model exhibiting a mean cross-validated AUROC of 0.7815 (95% CI: 0.657-0.886). The study demonstrated that the ensembled model, utilizing both clinical and HRV features, outperformed the model based solely on clinical parameters in eXtreme Gradient Boosting, Light GBM and Random Forest algorithms. The ensemble model exhibited the best performance in Light GBM algorithm (AUROC = 0.831; 95% CI 0.813-0.850), while HRV along exhibited an AUROC 0.809 (95% CI 0.792-0.825) in predicting neurological outcome in Gradient Boost algorithm. As to parameter analysis, Shapley's method identified Perinatal Index, Triangular Index, Mean Circular Variance of NN Intervals as key HRV features in predicting neurological outcomes. Conclusions: The integration of HRV features into machine learning models, combined with a rigorous internal cross-validation strategy, significantly enhanced prediction of neurological outcome in cardiac arrest. HRV features, alone or in combination with other clinical parameters, offered high interpretability and significant prognostic value in cardiac arrest. Future external validation in multicenter independent cohorts will further confirm the generalizability of these findings for clinical translation.