Structural heart disease detection using reconstructed compact electrocardiograms in the outpatient setting
T Paquaij, D Ahmetagic, B K O Arends, T P Mast, R R Van De Leur, P Van Der Harst, R Van EsAbstract
Background/Introduction
Heart failure is a progressive condition for which early detection improves clinical outcomes. Screening outside hospital settings remains challenging, as indicators of structural heart disease (SHD) are usually identified through transthoracic echocardiograms (TTEs). Recent studies show that neural networks trained on paired electrocardiogram-echocardiogram (ECG–TTE) data can detect SHD from electrocardiograms (ECGs), enabling the deferral of low-yield TTEs. In outpatient care, ECGs are often stored in compact formats (e.g. "3x4", "3x4+1" or "6x2") that lack continuous 10-second traces for all leads. Incomplete temporal information, particularly during irregular rhythms such as premature ventricular or atrial contractions, may reduce diagnostic performance.
Purpose
This study evaluates reconstruction of full 10-second ECGs from compact formats and its impact on SHD detection.
Methods
We trained a recurrent inference machine, ECG-RISE, to reconstruct continuous 10-second ECGs with randomly blanked segments. The training dataset included 95,831 physician-labelled ECGs encompassing a wide variety of conduction, ischaemic, and rhythm abnormalities. We performed SHD detection using ECGFounder, which we fine-tuned on 80,993 ECG–TTE pairs for classification across nine classes and a composite label. We compared the composite SHD classification performance of the original ECGs and reconstructed 3x4+1 ECGs, both as median beats.
Results
Evaluation was conducted in a multi-centre outpatient cohort of newly referred patients from one large and one academic dutch hospital (n = 4,036, 51% male, median age of 64 years). The reconstructed ECG median beat achieved an RMSE of 51 µV (IQR: 41.05-64.31) and a normalised corss-correlation coefficient of 0.96 (IQR: 0.94-0.98). For the composite SHD class (prevalence: 18.1%), original ECGs achieved an area under the receiver operator characteristic (AUROC) of 0.83 (95% CI: 0.81–0.85) and an area under the precision-recall curve (AUPRC) of 0.59 (95% CI: 0.57–0.63). Reconstructed ECGs yielded performance scores of 0.82 (95% CI: 0.80–0.84) and 0.57 (95% CI: 0.55–0.60), respectively. DeLong's testing demonstrated statistically unsignificant for AUROC (p = 0.788) and stratified bootstrap test demonstrated statistically significant for AUPRC (p = 0.006). Net reclassification index (NRI) showed -11.11% of reclassification based on the screening SHD threshold.
Conclusion(s)
Reconstruction of ECGs from compact formats resulted in minimal alterations on median beat morphology and preserved global model performance (AUROC). However, patient-level evaluation (.e.g. AUPRC and NRI) indicate adverse effect on individual predictions. These findings suggest that small reconstruction errors can have influence on medical decision making while not being captured by global metrics. Further research is needed to asses clinical implication of ECG reconstruction on SHD detection.For image description, please refer to the figure legend and surrounding text.