Basic/Translational Science -> Computer Modeling/Simulation D-PO02 - Poster Session II (ID 47) Poster

D-PO02-028 - A Novel Artificial Intelligence Model For Predicting Accessory Pathway Localization With Multimodality Image (ID 987)

 M. Nishimori: Nothing relevant to disclose.


Background: Catheter ablation is an established treatment for WPW syndrome, and it is important to predict localization of the accessory pathway (AP) in developing a rational treatment strategy. Several algorithms have been proposed to predict an AP from 12-lead electrocardiogram (ECG). However, it is difficult to predict the AP precisely because individual heart sizes and anatomical axes are individually different. In addition, we often face difficulties in judgement from ambiguous delta waveform and QRS polarity.
Objective: Regarding these problems, we have established a new predictive model using deep learning method.
Methods: ECG dataset of 88 WPW patients were divided into training set and test set, and training set were fitted to a 22-layer convolutional neural network to detect AP locations. Location labels were divided into three groups (left free wall, right free wall, and septal). To prevent overfitting, data augmentation and transfer learning were performed from 549 ECGs that could be used as open access resources. Moreover, to take into account for the individual differences in heart structure, chest X-ray image data was concatenated to ECG data and the accuracy of these models were compared.
Results: The maximum accuracy of the ECG-only model was 80.4%, and the accuracy of the model with ECG and XP data was improved by 4.0% (84.4%) compared to the previous model. (Shown in the Figure using smoothing method)
Conclusion: We thus developed a multi-modality machine learning model, which can be the novel method in establishing medical artificial intelligence model.