“What was AI thinking?": Explainable deep learning in reading of 12-lead ECGs for detecting V1 and V2 electrode misplacement

RR Bond, Khaled Rjoob, D Finlay, V. E. McGilligan, Stephen, J Leslie, D Guldenring, Ali Rababah, Aleeha Iftikhar, Charles Knoery, Anne McShane, Aaron Peace

Research output: Contribution to conferencePosterpeer-review

Abstract

Background: ECG data quality can be affected by lead misplacement which can affect clinical decisions. V1 and V2 are commonly misplaced in the superior direction from their correct position, which can mimic or conceal abnormalities. The aim of the current study is to use artificial intelligence (AI) in the form of deep learning to detect V1 and V2 lead misplacement to enhance ECG data quality and to make the black box decisions of AI systems more transparent
by providing AI attention maps.

Methods: V1 and V2 signals were collected from 453 patients (normal n=151, Left Ventricular Hypertrophy (LVH) n=151, Myocardial Infarction n=151) and extracted using a high-resolution body surface potential maps (BSPM) and converted into RGB images. A deep convolutional neural network (CNN) with 68 layers was developed and trained to classify the ECG images of V1 and V2 into correct and incorrect placement. An attention map was generated and analysed
for each ECG image in the last convolution layer to show the most important features (see figure 1) that have been selected by the CNN. CNN has been trained on 67% of the data and tested on 33%.

Results: Using CNN with 68 layers, the accuracy of detecting lead misplacement was 92.6% (TN=291/300, TP=265/300, FP=9/300, FN=35/300). Based on attention maps, P waves (56%), T waves (55%) and R (48%) waves contributed the most to the predicted classes correct and incorrect (see table 1). The S wave was not considered important in most cases in detecting correct V1 and V2 placement. The other features, including the PR interval, Q wave and J point
contributed 29%, 17% and 27% respectively to the predicted classes correct and incorrect.

Conclusions: Deep CNN achieved a high accuracy (92.6%) to detect V1 and V2 lead misplacement, whilst increasing the transparency of the algorithmic decision making. Attention maps demonstrate what the algorithm ’looked at’ prior to making it’s decision, which also
elucidate areas of the ECG that are most important in detecting lead misplacement. Physicians can use the attention map to calibrate their trust with the machine and to consider the machine’s attention (a proxy for machine rationale). According to the generated attention maps, the P waves, T waves and R waves were considered the most important features, while the S wave was considered as the least important feature. Whilst the other features PR interval, Q wave
and J point are considered as mid-level features.

Original languageEnglish
Publication statusPublished - 28 Apr 2021
Event45th International Society for Computerized Electrocardiology -
Duration: 28 Apr 20211 May 2021
Conference number: 45th
https://cdn.ymaws.com/www.isce.org/resource/resmgr/hauck/2021_conference/isce_abstract_booklet_4.30.2.pdf

Conference

Conference45th International Society for Computerized Electrocardiology
Abbreviated titleISCE
Period28/04/211/05/21
Internet address

Keywords

  • deep learning
  • explainable AI

Fingerprint

Dive into the research topics of '“What was AI thinking?": Explainable deep learning in reading of 12-lead ECGs for detecting V1 and V2 electrode misplacement'. Together they form a unique fingerprint.

Cite this