Survival prediction modelling in patients with acute ST-segment elevation myocardial infarction with LASSO regression and explainable machine learning.

BACKGROUND

Acute ST-segment elevation myocardial infarction (STEMI) is a cardiovascular emergency that is associated with a high risk of death. In this study, we developed explainable machine learning models to predict the overall survival (OS) of STEMI patients to help improve prognosis and increase survival.

METHODS

After applying the inclusion and exclusion criteria, we selected 893 patients who underwent emergency coronary angiography and percutaneous coronary intervention (PCI) for STEMI at the First Affiliated Hospital of Kunming Medical University. The best predictor variables were screened by least absolute shrinkage and selection operator (LASSO) regression. These variables were used to construct Cox proportional hazards regression (coxph) and random survival forest (rfsrc) models. Three criteria (C-index, Brier score, and C/D AUC) were utilised to compare the performance of the two models. Then, by applying the time-dependent variable importance and the partial dependence survival profile, a global explanation of the entire cohort was conducted. Finally, local explanations for individual patients were performed with the SurvSHAP(t) and SurvLIME plots and the ceteris paribus survival profile.

RESULTS

Combining the results of the comparison of the three criteria, the performance of the rfsrc model was shown to be superior to that of the coxph model. LASSO regression was used to screen 11 predictor variables, such as diastolic blood pressure (DBP), Killip class, hyperlipidaemia, global registry of acute coronary events (GRACE) Score, creatine kinase isoenzyme-MB, myoglobin, white blood cells, monocytes, thrombin time, globulin (GLB), and conjugated bilirubin. The global explanation of the whole cohort revealed that DBP, GRACE Score, myoglobin, and monocytes had a significant effect on the OS of STEMI patients in the coxph model and that DBP, GRACE Score, and GLB were the variables that significantly affected the OS of STEMI patients in the rfsrc model. Incorporating a single patient into the model can yield a local explanation of each patient, thus guiding clinicians in developing precision treatments.

CONCLUSION

The rfsrc model outperformed the coxph model in terms of predictive performance. Clinicians can use these predictive models to understand the major risk factors for each STEMI patient and thus develop more individualised and precise treatment strategies.