Table 1 Summary of methodologies, key features, and limitations in ECG based AI models.

Raymond et al.¹³

PTB-XL, CPSC-2018

VGGNet-16

Utilized identical way of transfer learning with ImageNet weights.

\(\bullet\) Predictions were based solely on 12-lead ECG.\(\bullet\) Model complexity due to high parameter count.

Eleyan A. et al.¹⁶

MIT-BIH, BIDMC

CNN-LSTM

A hybrid model incorporating Fourier transformation was used for heartbeat classification.

\(\bullet\) Increased computational complexity due to hybrid architecture. \(\bullet\) Single-lead ECG (SL-ECG) was not considered for predictions.

Annisa D. et al.¹⁷

Lobachevsky University database

CNN-LSTM

CNN was employed for feature extraction, while LSTM handled classification in a hybrid approach.

\(\bullet\) Computational overhead remains a concern. \(\bullet\) Training and testing were performed on identical datasets, limiting SL-ECG prediction.

Xiangyu Z. et al.¹⁸

CinC-2017

TCN, ResNet

Temporal Convolutional Network (TCN) was evaluated for time-series ECG analysis.

\(\bullet\) ResNet, used within the model, adds architectural complexity. \(\bullet\) Training and testing relied on the same SL-ECG dataset, demanding high SL-ECG data acquisition.

Jiwoong K. et al.²⁰

Private dataset

RNN, LSTM, ResNet-50

Self-collected SL-ECG dataset was used for model development.

\(\bullet\) High computational demands for the considered models. \(\bullet\) Training and testing were performed exclusively on SL-ECG data only. \(\bullet\) New data collection is resource-intensive and time-consuming.

Attia Z. et al.²³

Self Collection

CNN

A distinct heart condition was predicted using both 12-lead and SL-ECG data.

\(\bullet\) Lack of evaluation on widely used wearable SL-ECG devices. \(\bullet\) Insufficient analysis of differences between 12-lead and SL-ECG data discrepancies. \(\bullet\) Model sustainability regarding artifact removal and segmentation are lacking.