Treatment of individual predictors with neural network algorithms improves Global Registry of Acute Coronary Events score discrimination

Abstract Objective: The aim of this study was to develop, train, and test different neural network (NN) algorithm-based models to improve the Global Registry of Acute Coronary Events (GRACE) score performance to predict in-hospital mortality after an acute coronary syndrome. Methods: We analyzed a prospective database, including 40 admission variables of 1255 patients admitted with the acute coronary syndrome in a community hospital. Individual predictors included in GRACE score were used to train and test three NN algorithm-based models (guided models), namely: one- and two-hidden layer multilayer perceptron and a radial basis function network. Three extra NNs were built using the 40 admission variables of the entire database (unguided models). Expected mortality according to GRACE score was calculated using the logistic regression equation. Results: In terms of receiver operating characteristic area and negative predictive value (NPV), almost all NN algorithms outperformed logistic regression. Only radial basis function models obtained a better accuracy level based on NPV improvement, at the expense of positive predictive value (PPV) reduction. The independent normalized importance of variables for the best unguided NN was: creatinine 100%, Killip class 61%, ejection fraction 52%, age 44%, maximum creatine-kinase level 41%, glycemia 40%, left bundle branch block 35%, and weight 33%, among the top 8 predictors. Conclusions: Treatment of individual predictors of GRACE score with NN algorithms improved accuracy and discrimination power in all models with respect to the traditional logistic regression approach; nevertheless, PPV was only marginally enhanced. Unguided variable selection would be able to achieve better results in PPV terms.

[1]  M. Zdravković,et al.  Comparison of RISK-PCI, GRACE, TIMI risk scores for prediction of major adverse cardiac events in patients with acute coronary syndrome , 2017, Croatian medical journal.

[2]  Richard Segal,et al.  Risk prediction model for in‐hospital mortality in women with ST‐elevation myocardial infarction: A machine learning approach , 2017, Heart & lung : the journal of critical care.

[3]  Payam Amini,et al.  Prediction of Kidney Graft Rejection Using Artificial Neural Network , 2017, Healthcare informatics research.

[4]  Collin M. Stultz,et al.  Machine Learning Improves Risk Stratification After Acute Coronary Syndrome , 2017, Scientific Reports.

[5]  John Wallert,et al.  Predicting two-year survival versus non-survival after first myocardial infarction using machine learning and Swedish national register data , 2017, BMC Medical Informatics and Decision Making.

[6]  A. Hoes,et al.  Comparison of the GRACE, HEART and TIMI score to predict major adverse cardiac events in chest pain patients at the emergency department. , 2017, International journal of cardiology.

[7]  Collin M. Stultz,et al.  Beatquency domain and machine learning improve prediction of cardiovascular death after acute coronary syndrome , 2016, Scientific Reports.

[8]  Soleiman Kheiri,et al.  A Hybrid ANN-GA Model to Prediction of Bivariate Binary Responses: Application to Joint Prediction of Occurrence of Heart Block and Death in Patients with Myocardial Infarction , 2016, Journal of research in health sciences.

[9]  Imad H. Elhajj,et al.  Artificial neural network modeling enhances risk stratification and can reduce downstream testing for patients with suspected acute coronary syndromes, negative cardiac biomarkers, and normal ECGs , 2016, The International Journal of Cardiovascular Imaging.

[10]  Tao Zhang,et al.  Risk stratification and prognostic value of grace and timi risk scores for female patients with non-st segment elevation acute coronary syndrome. , 2015, International journal of clinical and experimental medicine.

[11]  Giuseppe Biondi-Zoccai,et al.  TIMI, GRACE and alternative risk scores in Acute Coronary Syndromes: a meta-analysis of 40 derivation studies on 216,552 patients and of 42 validation studies on 31,625 patients. , 2012, Contemporary clinical trials.

[12]  G. Toffolo,et al.  Artificial neural networks and robust Bayesian classifiers for risk stratification following uncomplicated myocardial infarction. , 2005, International journal of cardiology.

[13]  E R Bates,et al.  Comparison of artificial neural networks with logistic regression in prediction of in-hospital death after percutaneous transluminal coronary angioplasty. , 2000, American heart journal.

[14]  Geoffrey E. Hinton,et al.  A comparison of statistical learning methods on the Gusto database. , 1998, Statistics in medicine.