Integration of clinical and imaging data to predict the presence of coronary artery disease using neural networks

Abstract

Background: This study tests the hypothesis that a computer-based artificial intelligence method capable of integrating clinical, exercise and imaging data can determine the presence or absence of coronary artery disease (CAD) with predictive accuracy equal to that of standard, expert reader clinical interpretation using imaging, clinical and stress test data. Methods: Patients (Pts; n=102) underwent rest, stress (1 day protocol) gated SPECT technecium-99m-MIBI myocardial perfusion imaging (MPI) and were selected based on availability of high quality MPI and coronary angiography done an average of 60 days after the stress test (range 0–257). All Pts had standard Bruce protocol. Objective parameters (total =20) describing each Pt's stress responses (eg; ST segments, HR, BP, symptoms), clinical status (eg; gender, BMI) and image parameters (eg; SSS,SDS, LV mass, LVEF by Siemens 4DM program;) were used to train an artificial neural network (ANN). Pt's with prior myocardial infarction (N=19) were analyzed separately and since results were unchanged included in the totals. Classification accuracy (CAD, yes vs no) of ANN was compared with that of expert-readers. Extent and location of CAD by ANN also was compared to that of expert readers. A more limited ANN using imaging data only (ANNi) was compared to ANN using all imaging, clinical and stress test parameters (ANNt). Results: Pts were 65+11 yrs, 68% male and 75% had at least one 70% stenosis and so defined as CAD present. ANNt had sensitivity 88% and specificity 65% for CAD vs 80% and 69%, respectively by expert readers (P=ns for sens and spec). Predictive accuracy (PA= TP+TN/all Pts) was comparable (ANNt=81% v experts =77%, p=0.07). However, PA of ANNt (81%) was superior to ANNi (68%, z=2.26, P<0.05) for CAD diagnosis. ANNt and humans were comparable for detecting extent and location of CAD (83,93,92% of 1,2, 3 VD vs 78,82,83%, respectively). Conclusion: The data confirm the hypothesis that an ANN, with training using a probabilistic algorithm, is capable of integrating imaging, clinical and stress test data to predict CAD with accuracy equal to expert humans and superior to images only.