Fig 1.
Comparison of the standard scientific method and the deep–learning based approach.
Starting from a question/problem to be solved, the standard scientific method is based on a thorough literature research resulting in the construction of a hypothesis. Conversely, the deep–learning based approach uses already available data to train a neural network to solve the defined task as good as possible. Assessment of its performance requires comparison with previously established methods. The best performing networks are then analyzed by class activation mapping to visualize the underlying morphology triggering the network. Based on these visualizations, a hypothesis is constructed and tested in an appropriate experiment. Light–grey: shared processes; dark–grey: standard scientific method; white: deep–learning based, hypothesis–generating approach.
Fig 2.
Input data for the neural networks.
The input data were taken from the original study analysis and converted into jpgs with a predefined image size (224x224x3 pixels). Each image contained either the illustration of a non–ischemic (i.e., recorded directly before the coronary balloon occlusion) or an ischemic (i.e., recorded at the end of the balloon occlusion) intracoronary ECG as well as the corresponding label (non–ischemic respectively ischemic). In this example, both icECGs are from the same vessel (left anterior descending coronary artery) from the same patient. IcECG ST–segment shift was 0.056mV respectively 0.858mV.
Table 1.
Patient characteristics.
Table 2.
ST–segment shift and target vessel distribution.
Fig 3.
IcECG ST–segment shift grouped according to the state of absent or present coronary artery balloon occlusion.
Combination of the validation and the examination data (n = 225) was used for the performance analysis. Black circles: Non–ischemic prediction of ResNet5, black crosses: Ischemic prediction of ResNet5. Red signals: wrong predictions of ResNet5. Error bars indicate mean values and SD.
Fig 4.
Nonparametric receiver–operating characteristic curve of the icECG ST–segment shift and the network predictions using coronary artery patency or occlusion as dichotomic reference for absent or present myocardial ischemia.
Of note, network prediction provides a dichotomous output (non–ischemic respectively ischemic), resulting in a triangular ROC–curve. Hence, there is only one combination of sensitivity and specificity possible for each CNN. Dashed black line = reference line.
Fig 5.
Visualization of network activation patterns of the three best performing CNN.
Red regions contributed most to the network class prediction. ResNet5 bases its prediction on the area under the ST–segment and the T–wave. GoogLeNet10 is activated by the QRS–complex and the J–point for the non–ischemic state, and by the ST–segment and the T–wave for the ischemic state. ResNet6 bases its prediction on the ST–segment for the non–ischemic state, and on the end of the T–wave for the ischemic state. Please note the rather uncertain prediction of ResNet6 on the ischemic ECG.
Table 3.
Prediction and performance of the best ten trained networks.