A new study led by Dr Reza Argha introduces a novel deep learning framework designed to improve the reliability and generalisability of electrocardiogram (ECG) classification across diverse patient populations.
A new study led by Dr Reza Argha introduces a novel deep learning framework designed to improve the reliability and generalisability of electrocardiogram (ECG) classification across diverse patient populations.
The ECG is a cornerstone of cardiovascular diagnosis, widely used to detect arrhythmias and other cardiac abnormalities. However, ECG signals can vary significantly depending on patient characteristics, device type, recording environment, and clinical setting – a critical limitation for real-world and telehealth deployment.
In response to this challenge, Institute member Dr Argha and colleagues developed ECG-Adapt, a new artificial intelligence (AI) framework designed to help ECG models perform reliably when applied to new populations or clinical settings. Instead of assuming that data collected in one hospital or study will look the same as data collected elsewhere, the framework actively learns how to adjust to differences in patient characteristics, devices, and recording conditions. When applied to a new dataset without pre-existing diagnostic labels, the model gradually refines its predictions by checking and correcting its own early assumptions – preventing small early errors from compounding and reducing the risk of inaccurate classifications.
The study demonstrated meaningful improvements when models were transferred across datasets. For single-lead ECG tasks, ECG-Adapt improved overall classification performance by up to 8 per cent compared to directly applying an existing model. Similar gains were seen in 12-lead ECG tasks, with the framework outperforming several current approaches while remaining computationally efficient.
“Ensuring that AI models remain accurate when deployed across different hospitals, devices, and patient groups is essential for safe clinical adoption,” says lead author Dr Reza Argha. “ECG-Adapt demonstrates that we can significantly reduce performance degradation across datasets, improving reliability in telehealth environments.”
“As our population ages, we need AI tools that perform reliably across diverse and older patient groups. ECG-Adapt helps ensure that cardiac abnormalities can be detected accurately, even when models are deployed in new clinical or telehealth environments.”