Background: There is a substantial number of interrelated variables that have been associated to coronary artery disease (CAD) and to an increased risk of adverse cardiovascular outcomes. Currently, machine learning algorithms such as Artificial Neural Networks (ANN) constitute flexible models capable of characterizing complex relationships between numerous variables. ANN can be trained and applied to complex classification problems. On the other hand, PET perfusion imaging studies have demonstrated that a hampered myocardial perfusion reserve (MPR) (<2.0) conveys a significant risk for adverse cardiovascular outcomes. The present study generated and applied an ANN for the prediction of a PET-measured hampered MPR based on simple available predictors in patients with intermediate risk of CAD.
Methods: We included 1,241 patients with no previous MI or revascularization referred to 13N-ammonia PET for suspected ischemia. They were randomly divided into training, testing and holdout datasets (6:2:2). Demographic (sex, age, BMI and family history of CAD), clinical (chest complaints, diabetes, smoking, dyslipidemia, hypertension, rest heart rate [HR] and resting blood pressure [BP]) and complementary diagnostic data (Duke score, abnormal rest ECG, abnormal stress ECG, stress HR, % of max HR, stress BP, rest LVEF and stress LVEF) was retrieved. A multilayer perceptron analysis constructed the feed forward back propagation ANN with a hyperbolic tangent as the activating function. The network was trained and tested to classify the probability of cases of presenting a reduced MPR (<2.0). The ANN was then applied to the holdout dataset. The area under the curve (AUC) and normalised importance (NI) of the predictors were reported. Additionally, ensemble boosting of the ANN was performed to explore for accuracy improvement.
Results: The resulting ANN architecture used one hidden layer (with 6 nodes) and it showed a stable overall accuracy of 75% in the divided datasets, showing adequate generalisation. The ANN's AUC was 0.77 and the 5 most relevant predictors for network's construction were rest HR, age, stress, LVEF, BMI and stress HR (with a NI=100%, 63%, 41%, 32.5% and 31.3%). The ANN was generally more accurate when classifying a patient with a normal MPR than one with an abnormal MPR. Ensemble boosting, which iteratively created and improved the network classifiers achieved an increment of the overall accuracy to 84% when classifying new patients.
Conclusion: This study suggests that machine learning in the form of an ANN may be utilised to improve identification patients who will demonstrate a reduced MPR, based on simple and readily available clinical data, with a good overall accuracy (84%). This opens the possibility to improve the characterisation of patients at risk based on numerous and interrelated clinical variables. Further research into the application of ANNs with emerging and relevant predictors of cardiovascular risk is warranted.