Detecting cardiovascular diseases using ECG scans and explainable artificial intelligence

Cardiovascular diseases are the leading cause of mortality globally, requiring early and accurate detection through tools like electrocardiography. While artificial intelligence models have emerged to provide reproducible analysis of electrocardiogram printouts, their clinical deployment is hindered by a lack of transparency and sensitivity to real-world image variations such as discolorations, handwriting, or paper wrinkles. This study introduces an explainable artificial intelligence framework designed to quantify the stability of deep learning models and identify vulnerabilities in their behavior under controlled image perturbations.

We utilized a large-scale dataset of electrocardiogram printouts synthesized from the PTB-XL benchmark, creating both clean and contaminated versions featuring various image-level manipulations. Four deep learning architectures, including EfficientNet and InceptionNet, were trained and evaluated using different activation functions. The stability of these models was assessed using local interpretable model-agnostic explanations. We employed intersection over union metrics to measure the consistency of explanations across perturbations and extracted radiomic-like image features to quantitatively analyze the characteristics of the generated explanations.

Our experiments demonstrate that models trained on augmented datasets generalize better to perturbed data, with the best-performing model achieving an area under the receiver operating characteristic curve of 0.894 on the contaminated test set. Stability analysis showed that models trained on data containing perturbations achieved the highest average intersection over union of 0.399, indicating a more consistent focus on diagnostic features. Furthermore, radiomic-like features enabled the precise identification of the underlying deep learning model with an accuracy of up to 98%.

The proposed framework enables a comprehensive visual and quantitative evaluation of artificial intelligence stability in cardiovascular disease detection. By identifying how specific image manipulations affect model reliability, this approach can guide the development of more robust algorithms and targeted data augmentation strategies. To ensure full reproducibility and foster cross-domain collaboration, our tools and datasets are available at https://github.com/smile-research/xai-ecg.