AI Summary & Key Takeaways
AI InsightThe dataset used for model development, along with the executable Python scripts, is available for sharing.
📌 3 Key Takeaways
Abstract
Aims: Mortality rates remain high among patients admitted to the intensive care unit (ICU) following successful return of spontaneous circulation (ROSC) after cardiopulmonary resuscitation (CPR). Identifying risk factors specific to this patient group may directly inform clinical decision-making processes. This study aimed to identify the clinical and laboratory parameters associated with mortality in post-CPR ICU patients and to compare machine learning models developed using these parameterswith traditional statistical analyses. Methods: This retrospective study included a total of 82 patients treated in a tertiary-level ICU between 2020 and 2023. The post-CPR group (n=41) consisted of patients admitted to the ICU following effective CPR and ROSC, while the control group (n=41) included randomly selected patients with similar clinical characteristics who had not undergone CPR. Demographic data, clinical scores (APACHE II, SOFA, NUTRIC), laboratory values, and survival outcomes were recorded. Mortality prediction models were developed using the Random Forest algorithm applied to class-balanced datasets generated with the ADASYN method. Results: The post-CPR group had significantly higher scores and biomarker levels, including APACHE II, SOFA, and CRP, whereas albumin and GFR levels were notably lower. Both ICU and hospital mortality rates were significantly elevated in this group (75.6% and 80.5%, respectively; p<0.001). In general ICU mortality models developed using Random Forest, variables such as inotropic support, APACHE II, SOFA, and CRP emerged as prominent predictors, and the model demonstrated high predictive performance (AUC: 0.914). In the subgroup of post-CPR patients, factors such as thrombocyte count, mean platelet volume, and sex were found to be particularly influential in predicting mortality. Conclusion: Both traditional statistical analyses and machine learning models provide clinically meaningful results in predicting early mortality among post-CPR patients. In particular, the need for inotropic support and elevated inflammatory markers appear to be strong predictors of mortality. The high predictive performance of AI-supported models, even with small sample sizes, highlights their potential clinical utility, though prospective observational studies are needed to further validate these models. However, the limited cohort size and the absence of resuscitation-specific variables such as initial CPR rhythm and duration represent important limitations that should be addressed in future prospective studies. The dataset used for model development, along with the executable Python scripts, is available for sharing.
Özet
Aims: Mortality rates remain high among patients admitted to the intensive care unit (ICU) following successful return of spontaneous circulation (ROSC) after cardiopulmonary resuscitation (CPR). Identifying risk factors specific to this patient group may directly inform clinical decision-making processes. This study aimed to identify the clinical and laboratory parameters associated with mortality in post-CPR ICU patients and to compare machine learning models developed using these parameters with traditional statistical analyses. Methods: This retrospective study included a total of 82 patients treated in a tertiary-level ICU between 2020 and 2023. The post-CPR group (n=41) consisted of patients admitted to the ICU following effective CPR and ROSC, while the control group (n=41) included randomly selected patients with similar clinical characteristics who had not undergone CPR. Demographic data, clinical scores (APACHE II, SOFA, NUTRIC), laboratory values, and survival outcomes were recorded. Mortality prediction models were developed using the Random Forest algorithm applied to class-balanced datasets generated with the ADASYN method. Results: The post-CPR group had significantly higher scores and biomarker levels, including APACHE II, SOFA, and CRP, whereas albumin and GFR levels were notably lower. Both ICU and hospital mortality rates were significantly elevated in this group (75.6% and 80.5%, respectively; p<0.001). In general ICU mortality models developed using Random Forest, variables such as inotropic support, APACHE II, SOFA, and CRP emerged as prominent predictors, and the model demonstrated high predictive performance (AUC: 0.914). In the subgroup of post-CPR patients, factors such as thrombocyte count, mean platelet volume, and sex were found to be particularly influential in predicting mortality. Conclusion: Both traditional statistical analyses and machine learning models provide clinically meaningful results in predicting early mortality among post-CPR patients. In particular, the need for inotropic support and elevated inflammatory markers appear to be strong predictors of mortality. The high predictive performance of AI-supported models, even with small sample sizes, highlights their potential clinical utility, though prospective observational studies are needed to further validate these models. However, the limited cohort size and the absence of resuscitation-specific variables such as initial CPR rhythm and duration represent important limitations that should be addressed in future prospective studies. The dataset used for model development, along with the executable Python scripts, is available for sharing.