Clin Res Cardiol (2022). https://doi.org/10.1007/s00392-022-02002-5

Machine learning to optimize ECG leadwise evaluation for localization of myocardial infarction
E. Stützner1, N. Gumpfer1, S. Wegener2, J. Prim1, D. Grün2, J. Hanning1, M. Guckert1, T. Keller2
1Kompetenzzentrum für Informationstechnologie, Technische Hochschule Mittelhessen, Friedberg; 2Abteilung für innere Medizin I, Kardiologie, Justus-Liebig-Universität Giessen, Gießen;

Background and Aim:
The  utility  of  artificial  intelligence  (AI)  for  automated  evaluation  of  electrocardiogram (ECG) recordings has shown great potential. Identifying cardiac pathologies in the almost ubiquitously available, non-invasive ECG by an quick and cost-efficient AI-based system that is integrated into the daily workflow  may  become  a  significant  reliefs  for the treating physicians. One major limitation of such an approach is the black box nature of AI models.

Aim of this project is to combine different machine learning models to develop a diagnostic algorithm for identification and localization of myocardial infarction (MI) in 12-lead ECG data which allows the visualization of the decision process of the algorithm therefore improving trust.

Methods and Results:
ECG data used within this project stems from the PTB-XL database, a large publically available electrocardiography dataset containing n=21,837 clinical 12-lead ECG from n=18,885 patients. Of those, n=9,528 are defined as normal and n=5,486 are labeled with myocardial infarction. As first step, we  trained  twelve  AI  models (convolutional  neural  networks; CNNs),  one  for  each  lead  of  the  12-lead  ECG  recordings to identify anterior of inferior MI. As second step, these twelve distinct model predictions were then combined by application of a decision tree model as meta-learner. The advantage of this ensemble method is the gain of explainability by the hierarchically stacked individual lead information. The decision tree builds a hierarchy of the 12 input features, which can be visualized to explain the decisions of the ensemble, reflecting the relative importance of the different lead-based-probabilities. Stacked  models  for  identification of  anterior  and  inferior MI were developed.
The used ECG data was split for training and testing as follows: 50% of the Data was used to train and evaluate the lead specific CNNs. Of the remaining half, 80% were used to train the decision tree while the other 20% were used as test data. The presented results are based on this test dataset.

 

The two step machine learning approach is able to identify an anterior MI with a median accuracy of 90.8%, a median sensitivity of 83.1%, a median specificity of 94.58 % and a median area under the receiver operator characterist curve (AUROC) of 88.8 %. Here, the model mainly focuses on the anterior leads V2 and V3 (see Figure 1A presenting the model with a tree depth of 3).

Regarding differntiation of an inferior MI the model a median accuracy of 87.0 %. a  median sensitivity of 77.6 %, a median specificity of 91,9% and a median AUROC of 84.2%. IN the context of inferior MI identification the meta-learner identified the lead AVF as the most relevant one followed by the leads V3, V2,III, V5 and II (see Figure 2B).

Conclusion:
The  step-wise combination  of  convolutional  neural  networks  with  a  decision  tree using 12-lead ECG data yields competitive  results  for  identifying  localized myocardial  infarction  while  also  providing an  intuitive  source  of  visualization. This could facilitate development of trustworthy models for automated ECG evaluation as part of clinical decision support systems.








https://dgk.org/kongress_programme/jt2022/aV982.html