Exploratory development and internal validation of machine learning models for 30-day neurological outcome prediction after in-hospital cardiac arrest: a retrospective cohort study

Neurological outcomes after in-hospital cardiac arrest (IHCA) remain poor, with favorable recovery achieved in only 10%–25% of patients. Early prognostication is essential to guide treatment decisions and resource allocation; yet, accurate and interpretable predictive models remain scarce. This exploratory, hypothesis-generating study aimed to evaluate the discriminative performance of machine learning approaches for predicting 30-day neurological outcomes after IHCA using readily obtainable clinical variables. We retrospectively analyzed data from 124 IHCA patients (January 2020–April 2024). Given the limited sample size and class imbalance (favorable outcomes: 17.7%), we adopted bootstrap resampling with out-of-bag (OOB) validation (1,000 iterations). Features were selected via Recursive Feature Elimination (RFE) using AUC-ROC within each bootstrap iteration. The bootstrap training set was processed with the Synthetic Minority Over-sampling Technique (SMOTE) combined with Tomek links, while OOB samples retained their original class distribution. The optimal feature subset was determined by bootstrap stability analysis (selection frequency ≥ 80%). Nine algorithms were compared using OOB validation. To compare machine learning performance with established approaches, we additionally evaluated a conventional logistic regression model (LR-conv) based on the five selected predictors and the established Good Outcome Following Attempted Resuscitation (GO-FAR) score. XGBoost achieved the highest mean OOB AUC (0.81 ± 0.09) among the machine learning algorithms. LightGBM demonstrated comparable mean OOB AUC (0.80 ± 0.09), with acceptable calibration (Brier score 0.14 ± 0.04). Five predictors showed selection frequencies > 80%: resuscitation time (100.0%), albumin (99.0%), ALB/WBC ratio (96.7%), white blood cell count (90.7%), and age (90.3%).LR-conv achieved a mean OOB AUC-ROC of 0.70 ± 0.09 using the same five predictors. The GO-FAR score demonstrated a mean AUC of 0.73 ± 0.07 using bootstrap estimation. XGBoost and LightGBM demonstrated promising discriminative performance for 30-day neurological outcome prediction after IHCA in this small single-center cohort. These findings should be interpreted as hypothesis-generating, and external validation is required before clinical implementation. This study was not prospectively registered. A formal study protocol was not prepared prior to data analysis.