Superior machine learning model for post-cardiac arrest mortality prediction: a MIMIC-IV cohort study

Background: Post-cardiac arrest syndrome carries substantial mortality despite advances in resuscitation. We developed a machine learning model to predict 28-day mortality using comprehensive clinical data. Methods: We analyzed data from 1122 cardiac arrest patients in the MIMIC-IV database. After applying exclusion criteria, 853 patients with complete 28-day outcome data were included. We extracted 99 variables across six domains and compared five machine learning algorithms. Lactate clearance was calculated as: (Day 1 lactate - Day 3 lactate)/Day 1 lactate × 100%. Model performance was evaluated using AUC-ROC, calibration metrics, and SHAP analysis. Results: Among 853 patients (mean age 64.7 ± 16.3 years, 63.5% male), 327 (38.3%) died within 28 days. The XGBoost model achieved an AUC-ROC of 0.89 (95% CI: 0.86-0.92), outperforming APACHE III (AUC: 0.73). Lactate clearance rate emerged as the primary predictor (SHAP value: 0.24), followed by GCS score (0.21), SOFA score (0.18), age (0.16), and treatment intensity (0.14). Poor lactate clearance (0-25%) was associated with 67% mortality compared to 39% in moderate clearance (25-50%). Conclusions: Our machine learning model demonstrated superior accuracy for 28-day mortality prediction in post-cardiac arrest patients. Dynamic lactate clearance and neurological assessment provide actionable clinical insights for risk stratification.