Machine Learning Model for Predicting Risk of In-Hospital Mortality after Surgery in Congenital Heart Disease Patients

Interpretable machine learning for in-hospital mortality risk prediction in patients with ST-elevation myocardial infarction after percutaneous coronary interventions

BACKGROUND AND OBJECTIVE

Despite the constant improvement of coronary heart disease (CHD) diagnostics and treatment methods it remains one of the main causes of death in most countries around the world. And myocardial infarction with ST segment elevation on the electrocardiogram (STEMI) still is one of the most dangerous clinical variants of CHD. This study aims to develop an explainable machine learning model for in-hospital mortality (IHM) risk prediction in STEMI patients after myocardial revascularization by percutaneous coronary intervention (PCI).

METHODS

A single-center observational retrospective study was conducted, enrolling 4677 electronic medical records of patients with STEMI after PCI, which were analyzed using statistical analysis and machine learning methods. A pool of potential IHM predictors was identified, and prognostic models were developed and validated based on multivariate logistic regression, random forest, and stochastic gradient boosting methods at two stages of hospital treatment: during the initial physicians examination in the emergency department and immediately after PCI surgery. To explain the IHM prognosis, threshold values of IHM risk factors were determined using 3 grid search methods for optimal cut-off points, calculating centroids and SHapley Additive exPlanations (SHAP).

RESULTS

IHM prognostic models were developed using clinical and functional status data of STEMI patients during two stages of hospital treatment. The IHM prediction accuracy according to the first scenario was AUC = 0.85, and according to the second - AUC = 0.9. Predictors identified and validated in the models were converted into risk factors. Models whose parameters were risk factors demonstrated high forecast accuracy (AUC = 0.87), with the best model formed using the SHAP method.

CONCLUSIONS

For the forecast result interpretation risk factors obtained by categorizing continuous variables can be used by assessing the impact of the latter on the end point using the SHAP method.

Performance of the Models Predicting In-Hospital Mortality in Patients with ST-Segment Elevation Myocardial Infarction with Predictors in Categorical and Continuous Forms

The aim of the study is to assess the performance of predictive models developed on the basis of predictors in the continuous and categorical forms to predict the probability of in-hospital mortality (IHM) in patients with ST-segment elevation myocardial infarction (STEMI) after percutaneous coronary intervention (PCI).

Materials and Methods

A single-center retrospective study has been conducted, within the framework of which data from 4674 medical records of patients with STEMI after PCI, treated at the Regional Vascular Center of Vladivostok (Russia), have been analyzed. Two groups of patients were identified: group 1 consisted of 318 (6.8%) individuals who died in the hospital, group 2 included 4356 (93.2%) patients with a favorable outcome of treatment. IHM prognostic models were developed using multivariate logistic regression (MLR), random forest (RF), and stochastic gradient boosting (SGB). 6-metric qualities were used to evaluate the accuracy of the models. Threshold values of IHM predictors were determined using a grid search to find the optimal cut-off points, calculating centroids, and Shapley additive explanations. The latter helped evaluate the degree to which the model predictors influence the endpoint.

Results

Based on the results of the multi-stage analysis of indicators of clinical and functional status of the STEMI patients, new predictors of IHM have been identified and validated, complementing the factors of the GRACE scoring system, their categorization has been carried out and prognostic models with continuous and categorical variables based on MLR, RF, and SGB have been developed. These models had a high (AUC - 0.88 to 0.90) and comparable predictive accuracy, but their predictors differed in various degrees of influence on the endpoint. The comparative analysis has shown that the Shapley additive explanation method has advantages in categorizing predictors compared to other methods and allows for detailing the structure of their relationships with IHM.

Conclusion

The use of modern data mining methods, including machine learning algorithms, categorization of predictors, and assessment of the degree of their effect on the endpoint, makes it possible to develop predictive models possessing high accuracy and the properties of explanation of the generated conclusions.