Fusion-based QT interval estimation for enhanced long QT detection on a South African population using mobile ECG

Abstract Objective. This study compared QT interval estimation algorithms and their fusion on a South African population with a high prevalence of tuberculosis using a mobile electrocardiogram (ECG) to evaluate the precision of QT interval measurement and long QT detection. Approach. The study evaluated one open-source signal processing-based QT-estimation algorithm and four commercial machine learning-based algorithms (from AliveCor Inc., PulseAI Ltd., GE Healthcare Technologies Inc. and Safebeat Rx Inc.). Five algorithms were combined using two fusion models, median-based (MB) and linear regression-based (RB). These algorithms were validated on a South African Database (2050 2-lead ECGs) and tested on a separate database (692 patients with heart disease, where 50% had long-QT syndrome). Simultaneous expert overreads from 12-lead ECGs were used for reference. Main results. Both fusion algorithms achieved the highest overall performance and lowest QT estimation errors among all tested methods. On the validation set, the RB model showed a mean difference of 0.01 ± 17.00 ms, while MB showed −3.56 ± 17.55 ms. On the test set, the results were 0.14 ± 21.91 ms (RB) and −7.69 ± 21.17 ms (MB). For long QT detection, both algorithms ranked among the top performers, where RB reached the best area under the receiver-operating characteristic curve of 0.91 on the validation set and 0.96 on the test set. Significance. The study showed that an ensemble of independently developed QT estimation methods significantly enhanced the accuracy of automated QT analysis and long QT detection using a 2-lead mobile ECG. Algorithm fusion may reduce the likelihood of error and increase trust in automated medical diagnosis.