Abstract Body (Do not enter title and authors here): Background: AI algorithms for ECG images (AI-ECG) offer scalable screening and diagnosis of cardiac conditions by using images directly. However, the accuracy of predictions is highly dependent on data quality, with subjective human assessment of image quality causing harm during deployment. Therefore, we develop and validate an automated ECG quality adjudication system and evaluate model performance before and after its deployment (Fig A).
Research Question: Can a dedicated image quality assessment model (ECG-IQ) reliably identify low-quality ECG images and enhance AI-ECG diagnostic performance across diverse clinical settings?
Methods: We developed a convolutional neural network (ECG-IQ), designed to filter low-quality ECG images. The model was trained on a curated dataset of (i) clinician-annotated “poor quality” ECGs, (ii) synthetically generated low-quality ECGs, which were noised to simulate a range of common artifacts, such as baseline wander and powerline interference, across varying signal-to-noise ratios. We reserved 20% of the clinician-defined “poor quality” ECGs for internal validation. We examined the utility of ECG-IQ stratification by assessing the performance of previously developed and validated AI-ECG models for human detectable rhythm disorder atrial fibrillation, and invisible features of left ventricular systolic dysfunction, age, and sex, in multinational cohorts consisting of data from Yale New Haven Hospital, UK Biobank, and CODE-Brazil.
Results: In the Yale validation set (n=33,090, median age = 65, 49.6% female), ECG-IQ accurately identified real-world poor-quality ECGs (AUROC 0.93). Furthermore, a Grad-CAM analysis confirmed the model identified known artifact regions in low-quality ECGs (Fig B). When applied across multinational cohorts (n = 793,134, median age 57, 56.7% female), ECG-IQ stratification substantially improved the performance of multiple AI-ECG models. Specifically, in high-quality ECGs (as defined by ECG-IQ), the AUROC improved by an average of 32.8% for AF detection, 11.7% for LVSD, and 11.0% for sex classification, compared with low-quality ECGs. (Fig C). Similarly, median absolute error for age prediction decreased by an average of 27.0% in high-quality ECGs, compared with low-quality ECGs across all cohorts.
Conclusion: ECG-IQ reliably identifies low-quality ECGs, enabling the targeted deployment of AI-ECG models and improving their accuracy and safety in clinical use.