Optimal mean arterial pressure for favorable neurological outcomes in patients after cardiac arrest

Abstract

Background: Optimal mean arterial pressure (MAP) range after cardiac arrest remains uncertain. This study aimed to investigate the association between MAP and neurological outcomes during the early post-resuscitation period, with the goal of identifying optimal MAP range associated with favorable outcomes.

Methods: This retrospective observational study included 291 post-cardiac arrest patients treated at a tertiary care center. Five machine learning models to predict favorable neurological outcomes using hourly MAP measurements during the first 24 h after return of spontaneous circulation (ROSC) were compared and Random Forest model was selected due to its superior performance. Variable importance and Shapley Additive exPlanations (SHAP) were used to investigate the association between MAP and favorable neurological outcomes. SHAP dependence plots were used to identify optimal MAP ranges associated with favorable outcomes. In addition, individual-level predictions were interpreted using local interpretable model-agnostic explanations (LIME) and SHAP force plots.

Results: Machine learning analysis showed that MAP were associated with favorable neurological outcomes, with higher variable importance during the first 6 h after ROSC. SHAP analysis revealed an inverted U-shaped relationship between MAP and favorable neurological outcomes, with an optimal threshold of 79.56 mmHg (IQR: 73.70-82.54). This threshold remained consistent across both early (1-6 h: 79.26 mmHg) and later (7-24 h: 80.09 mmHg) hours. Individual-level explanations using SHAP and LIME highlighted that maintaining higher MAP during the early post-resuscitation period contributed positively to outcome predictions.

Conclusions: Machine learning analysis identified MAP as a major predictor of favorable neurological outcomes, with higher variable importance during the first 6 h after ROSC. MAP showed an inverted U-shaped relationship with favorable neurological outcomes, with an optimal threshold of approximately 80 mmHg.

Keywords: Cardiac arrest; Explainable machine learning; Mean arterial pressure; Neurological outcomes; Post-cardiac arrest care.

Fig. 1

Study population. KUED Korea University Emergency Department, CPR cardiopulmonary resuscitation, ROSC return of spontaneous circulation, ECLS extracorporeal life support, CPC cerebral performance category

Fig. 2

Variable importance rankings and SHAP summary plot of the Random Forest model predicting favorable neurological outcomes after cardiac arrest. A Bar plot indicating the variable importance of major predictors based on mean decrease in impurity. B SHAP summary plot illustrating the direction and magnitude of each predictor’s impact on the model output. MAP_#H mean arterial pressure recorded at #hour after return of spontaneous circulation, SHAP Shapley Additive exPlanations, MAP mean arterial pressure, CPR cardiopulmonary resuscitation, CA cardiac arrest

Fig. 3

SHAP dependence plots for mean arterial pressure at different timepoints. A–D SHAP dependence plots for MAP measured at 1H, 2H, 4H, and 6H. The x-axis represents MAP values (mmHg), the y-axis represents SHAP values, the orange curve is the LOWESS fit with its 95% confidence band (shaded). SHAP values greater than zero indicate an increased probability of a favorable neurological outcome. Red and blue markers denote the lower and upper MAP cutoff values identified for each timepoint. E Lower and upper MAP cutoff values during the first 24 h after return of spontaneous circulation. Asterisks (*) indicate timepoints, where an upper cutoff could not be derived. MAP_#H mean arterial pressure recorded at #hour after return of spontaneous circulation, SHAP Shapley Additive exPlanations, LOWESS LOcally Weighted Scatterplot Smoothing

Fig. 4

Patient-specific LIME and SHAP explanations for neurological outcome prediction. A Case 1—unfavorable neurological outcome (predicted probability = 0.20). B Case 2—favorable neurological outcome (predicted probability = 0.72). The upper panel is a LIME bar chart: features are ranked by influence, with red bars pushing the prediction toward a favorable outcome and blue bars pulling it toward an unfavorable one; the table to the right shows the patient’s raw values. The lower panel is a SHAP force plot: the gray vertical line marks the model’s baseline probability; red segments shift the prediction rightward (higher probability) and blue segments leftward (lower probability); segment width is proportional to each feature’s contribution. MAP_#H mean arterial pressure measured # h after return of spontaneous circulation (ROSC), VIS vasoactive-inotropic score, LIME local interpretable model-agnostic explanations, SHAP shapley additive explanations