Artificial intelligence (AI)-based evaluation of electrocardiograms (ECGs) have been used for detection of multiple cardiovascular diseases. Currently, there is a wide variety of AI methods used in clinical research. We evaluated the predictive ability of multiple approaches using neural networks, random forest, and support vector machines on structured ECG data regarding obstructive coronary artery disease (CAD) and long-term mortality in a cohort of patients undergoing conventional coronary angiography (CA). 

Our analysis is based on the cohort of the ECAD registry, including patients undergoing conventional CA at the West German Heart and Vascular Center between 2004 and 2019. Patients with a digitally available resting ECG within 90 days prior to coronary angiography, containing structured data on 648 characteristics for each ECG, were included. The overall cohort was divided into a learning cohort (60%), a validation cohort (20%) and a test cohort (20%). The incidence of death due to any cause was evaluated during follow-up. Logistic regression (LR) and linear discriminate analysis (LDA) were used as benchmarks. All neural nets were run 100x and averaged. An initial neural net (NN1) was improved by a sophisticated hyperparameter tuning, dropout layers and weight regularization (NN2). Feature reduction lead to NN3. Furthermore, other approaches like random forests (RF1), its adaption for the risk factors (RF2) and support vector machines (SVM) were calculated. All methods were compared by its mean area under the receiver operating curve (AUC) in the validation cohort.

 

We evaluated data from 7076 coronary angiographies. 2075 times (29.3 %), an obstructive CAD was discovered. During a median follow-up of 2.4 years (Q1: 0.8; Q3: 6.3), 1137 patients (16.1 %) died. A comparison of different calculatory approaches can be found in Table 1.  Overall, neural networks with feature selection led to highest AUC. 

 

Artificial intelligence algorithms can use secondary ECG data to hold up with cardiovascular risk factors in the detection of obstructive CAD and the prediction of mortality prior to coronary angiography.  In our study neural nets and random forests performed better then logistic regression, linear discriminant analysis and support vector machines. Dropout layers and weight regularization for neural nets, weighting the different independent variables for random factors and feature reduction in general were confirmed to increase the AUC.  

 

 

Table 1: Receiver operating characteristics for various AI methods for prediction of obstructive CAD and all-cause mortality

	LR	LDA	NN1	NN2	NN3	RF1	RF2	SVM
(A) Obstructive CAD
ECG variables alone	0.567	0.568	0.584	0.603	0.623	0.599	S	0.542
Traditional risk factors alone	0.576	0.576	0.586	0.590	S	0.568	S	0.536
ECG and risk factors	0.584	0.587	0.599	0.636	0.656	0.609	0.622	0.547
(B) All-cause mortality
ECG variables alone	0.590	0.618	0.627	0.659	0.707	0.659	S	0.641
Traditional risk factors alone	0.581	0.582	0.567	0.580	S	0.580	S	0.559
ECG and risk factors	0.600	0.627	0.610	0.663	0.711	0.660	0.663	0.651

S	Combination does not make sense
LR	logistic regression
LDA	linear discriminate analysis
NN1	a first trial of neural networks
NN2	neural networks with hyperparameter tuning and overfitting reduction
NN3	neural networks with feature reduction
RF1	random forests
RF2	random forest with increased probabilities for the risk factors
SVM	support vector machines