Interpretable machine learning model for predicting myocardial injury in intensive care unit patients using SHapley Additive exPlanations analysis

Abstract

ObjectiveThe identification of myocardial injury in the intensive care unit (ICU) has received little attention from researchers. Therefore, this retrospective cohort study aimed to develop a machine-learning model to predict the occurrence of myocardial injury in the ICU.MethodsBased on the Clinical Research Data Platform of Peking University People's Hospital, we enrolled adult, non-cardiac surgical, and non-obstetric patients who were admitted to the ICU between 2012 and 2022. Logistic regression, random forest, LASSO regression, support vector machine and extreme gradient boosting (XGBoost) models were developed to predict myocardial injury.ResultsData from 7453 non-cardiac surgery adult patients in ICU were collected in the derivation cohort (myocardial injury group: 2161 [29%], non-myocardial injury group: 5292 [71%]). Among the five models, the XGBoost model (area under the curve = 0.779; accuracy = 0.781) exhibited the best predictive performance for myocardial injury and the results were explained by the SHapley Additive exPlanations analysis. The top six features of the XGBoost model were maximal heart rate, respiratory rate, temperature, minimal heart rate, age and plasma transfusion.ConclusionThis machine-learning model, developed using the XGBoost algorithm, could be a valuable tool for clinical decision-making and detecting myocardial injury in the ICU.

Keywords: Intensive care unit; SHapley Additive exPlanations; machine learning; myocardial injury; non-cardiac surgery.

Figure 1.

Flowchart of screening. Flow chart of patients enrolled to build a predictive machine learning model for myocardial injury. ICU, intensive care unit.

Figure 2.

(a) Receiver operating characteristic curves of the five predictive models in the external validation cohort. XGB yielded the highest AUC for single-model prediction. (b) Multiple decision curve of five predictive models. The x-axis indicates the threshold probability for the myocardial injury outcome, and the y-axis indicates the net benefit.

Figure 3.

(a) A variable importance bee swarm plot was constructed using sHapley additive exPlanations (SHAP) summary tool for the myocardial injury prediction extreme gradient boosting model. According to the SHAP value of each feature, the maximal HR (h _max), maximal respiratory rate (rr_max), and age (yellow dots) were associated with a higher probability of myocardial injury (right side of the vertical dotted line). (b) Variable importance bar chart showing the mean absolute SHapley Additive exPlanations (SHAP) value of each feature and indicates its effect on the output of the extreme gradient boosting model.

Figure 4.

(a) Single-sample waterfall plot for patient ID 10 showing each step from the baseline value to the final predicted value, with each step corresponding to the contribution of a feature. Each level in the waterfall plot represents the contribution of a feature, which can be positive or negative. By observing the cumulative effects of these contributions, we can identify the features that have a significant impact on the final predicted value. For patient ID 10, the overall impact was negative. (b) Single-sample waterfall plot for Inpatient ID 1010, the overall impact is positive, in which the feature of maximal HR (hr_max) contributes the most to the predictive output. (c) Force plot for patient ID 10 shows the relationship between the feature values of each sample and the corresponding SHapley Additive exPlanations (SHAP) values. The features of patient ID 10 are represented in purple and indicate a negative probability of myocardial injury. (d) Force plot for patient ID 1010, in contrast to (c), the features of patient ID 1010 are represented in yellow and purple, where the purple part is overwhelmed by yellow, indicating a positive probability of myocardial injury.

Figure 5.

Partial dependence plot. Each dependence plot shows how the features affect each other and the output of the predictive model, where each dot represents one patient. SHapley Additive exPlanations (SHAP) values are represented on the y-axis. The x-axis represents the actual values of one feature, and the values of the other features are depicted by a colour gradient from yellow to purple in descending order. A SHAP value for specific features exceeding zero pushes the decision toward myocardial injury.