New score for evaluation of intrahospital outcome of patients with acute myocardial infarction

Introduction: Risk stratification has become an integral component of modern treatment in clinical practice. Procedural processes during primary percutaneous coronary intervention (pPCI) as well as knowledge about the distribution and types of lesions in coronary arteries are of great importance, and a final risk evaluation is recommended directly after the pPCI. Methods of data mining allow finding hidden patterns in data, but also development of modern predictive models. Purpose: To create and test a simple, practical and usable predictive model in daily practice for the assessment of intrahospital treatment outcome of patients with acute myocardial infarction with STsegment elevation (STEMI) treated with pPCI. Methods: Presented research is unicentric, retrospective but also prospective study. Retrospective study included 1495 patients with STEMI admitted to our hospital and treated with pPCI during the period from December 2008 to December 2011. Each patient was initially described by 629 attributes (demographic characteristics, data from history and clinical findings, biochemical parameters of blood tests on admission, the echocardiographic parameters, angiographic and procedural details and admission diagnosis codes). For model development, an open source software solution Weka was used. During the evaluation of different algorithms, algorithm that gives the best results in terms of accuracy and ROC parameter was chosen. As part of the retrospective study, in order to assess the models performance, tenfold crossvalidation on the entire data set was used. Additional validation of the developed predictive model was conducted in a prospective study, on a sample of 400 patients with STEMI, treated with pPCI in 2015. In addition, GRACE risk score was calculated for the prospective study patients and comparison with the developed model has been performed. Results: Alternative decision tree (ADTree) was selected as best performing algorithm. Cost sensitive classification was used as an additional methodology to enhance accuracy. ADTree selected eight key parameters that most influence the outcome of intrahospital treatment: systolic blood pressure on admission, left ventricular ejection fraction, stroke volume of the left ventricle, troponin, creatine phosphokinase, total bilirubin, T wave shape and the result of the intervention. The performance of the developed model are: the accuracy of the prediction is 93.17%, ROC 0.94. The developed model kept its performance in prospective validation: accuracy of prediction 90.75%, ROC 0.93. Widely used GRACE score achieved ROC = 0.86 in the prospective study patients, indicating that developed predictive model is superior. Conclusion: Developed predictive model is simple and reliable. Its implementation in everyday clinical practice, would allow clinicians to distinguish highrisk patients after reperfusion treatment, and then to intensify treatment and clinical followup.