Predicting 30-day survival after in-hospital cardiac arrest: a nationwide cohort study using machine learning and SHAP analysis

利用机器学习和SHAP分析预测院内心脏骤停后30天生存率:一项全国性队列研究

阅读:1

Abstract

OBJECTIVE: In-hospital cardiac arrest (IHCA) presents a critical challenge with low survival rates and limited prediction tools. Despite advances in resuscitation, predicting 30-day survival remains difficult, and current methods lack interpretability for timely decision-making. This study developed a machine learning (ML) model to predict 30-day survival after IHCA, using peri-arrest variables available on the rescue team's arrival, while ensuring a balance between predictive accuracy and clinical interpretability through Shapley Additive Explanations (SHAP). DESIGN: A nationwide, registry-based observational study. SETTING: Data were sourced from the Swedish Cardiopulmonary Resuscitation Registry (2010-2020), merged with the Patient Registry. PARTICIPANTS: We analysed 25 905 IHCA cases with attempted resuscitation, of which 8166 patients survived for 30 days. OUTCOME MEASURE AND ANALYSIS: 30-day survival after IHCA was the outcome measure. An ML model was developed using fivefold cross-validation. Key predictors were identified through in-built variable importance and validated using SHAP. Model performance was evaluated with metrics such as area under the receiver operating characteristics (AUROC), calibration, sensitivity, specificity, false negative rate (FNR) and F-score. RESULTS: The CatBoost model achieved an AUROC of 0.9136 (95% CI 0.9075 to 0.9191) with all features, and 0.9034 (95% CI 0.8955 to 0.9037) with the top 15 features, along with Brier scores of 0.1028 and 0.1103, respectively. Performance plateaued after including the top 15 predictors, with few key variables, such as epinephrine administration, age, initial rhythm, ROSC within 15 min, breathing on rescue team arrival and witnessed cardiac arrest, being most influential. The model showed strong calibration for patients with low predicted survival probabilities and demonstrated high sensitivity with a low FNR across relevant survival thresholds. CONCLUSION: The CatBoost model provides an effective and interpretable tool for predicting 30-day survival after IHCA. Key predictors such as epinephrine administration, age and initial rhythm inform clinical decision-making. This model has strong clinical utility and can be externally validated via the open-access Application Programming Interface (API) at www.gocares.se.

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。