Deep Learning Model to Predict Mortality in Patients Discharged After ST-Segment Elevation Myocardial Infarction Treated by Percutaneous Coronary Intervention (From the PL-ACS Registry)

Background

The realtive incidence of ST-elevation myocardial infarction (STEMI) and in-hospital mortality are decreasing. Further improvement of survival may be achieved by intesing care in individuals with an increased risk of unfavorable events after discharge. To select such a threatened population, various statistical models may be implemented, including regression or artificial inteligence. One of the novel approach is deep learning.

Our aim was to built and compare two prediciton models using logistic regression and deep learning methods.

Methods

The study included all of the STEMI patients who were treated with primary percutaneus coronary intervention (pPCI) and discharged alive who were registered in the prospective Polish Registry of Acute Coronary Syndromes (PL-ACS) from January 2009 to December 2014. The patients with prior myocardial infarction or those treated with thrombolysis were excluded. The 12-month mortality was obtained from a government database. Both models included the same 104 variables and were biult and tested/validated in the same patients’ subgroups.

Results

Among 62,586 patients discharged alive after STEMI, 4,460 (7.1%) died in 12-month follow-up. The logistic regression model included 42 variables with area under ROC curve (AUC) 0.814 for validation subgroup. In deep learning approach, neural network with 30 neurons in hidden layer and more than 5.400 weights was built with AUC 0.818 for validation subgroup.

Conclusions

The predictive value of deep learning approach in STEMI patietnts was at least as good as the model based on regression model.