AI-powered 12-lead ECG analysis for detecting high-risk PFO or symptomatic ASD
P Trochimiuk, W Ksiazek, M Gandor, J Kwiecinski, K A Sadowski, M Kania, K Kryczka, E Kaczmarska, Z Dzielinska, L Kalinczuk, P Plawiak, M DemkowAbstract
Background
Patent foramen ovale (PFO) and atrial septal defect (ASD) are common in patients with cryptogenic stroke or right-sided cardiac enlargement. Their diagnosis relies on transesophageal echocardiography (TEE), which is invasive and resource-intensive. Artificial intelligence applied to routine electrocardiography (ECG) may enable non-invasive detection of subtle electrical signatures of PFO/ASD.
Aims
To develop and evaluate AI-ECG models for identifying high-risk PFO or symptomatic ASD from resting digital ECGs, and to compare prior single-lead with newer full 12-lead architectures.
Methods
In this retrospective proof-of-concept study, resting 10-second ECG XMLs were analyzed using a balanced 80/10/10 split (training: 487 PFO/ASD and 487 controls; validation: 61 and 61; test: 61 and 61). Prior single-lead convolutional neural networks using lead II or V1 were compared with new 12-lead models trained on ECG signals with wavelet transform applied, including simple 1D and 2D architectures and an adapted residual-attention architecture. Each configuration was trained 30 times. Performance was assessed using accuracy, F1-score, sensitivity, specificity, predictive values, and area under the receiver-operating-characteristic curve (AUC).
Results
The best earlier single-lead model achieved accuracy 0.7787, sensitivity 0.8361, specificity 0.7213, and AUC 0.7845. Simple 12-lead CNNs yielded only modest gains, with mean accuracy 0.7036-0.7194 and mean AUC 0.7842-0.8046. In contrast, the adapted 12-lead residual-attention model substantially improved performance, reaching mean accuracy 0.8063 +/- 0.0258, F1-score 0.8178 +/- 0.0306, sensitivity 0.8765 +/- 0.0799, specificity 0.7361 +/- 0.0709, and AUC 0.8879 +/- 0.0223. The best test model achieved accuracy 0.8525, sensitivity 0.9016, specificity 0.8033, PPV 0.8209, NPV 0.8909, and AUC 0.8775.
Conclusions
Using all 12 leads alone did not guarantee better performance; the major gain emerged when multi-lead input was paired with a more expressive residual-attention architecture. These findings support further development of AI-enabled resting ECG as a scalable, non-invasive screening tool to guide referral for confirmatory TEE in patients at risk for PFO/ASD. External validation is required before clinical implementation.AI-ECG for PFO/ASD Performance by model