Cardiovascular diseases are the leading cause of death globally, with over 30% of all fatalities annually. Early and accurate detection is crucial to tackle heart failure through timely intervention. Electrocardiogram (ECG) signals are vital in diagnosing a wide range of cardiac conditions. In recent years, Artificial Intelligence (AI)-based models have shown to outperform humans in ECG interpretation. However, training these models often requires large labeled datasets, significant training time, and computational resources. In addition, the lack of datasets and training codes used in many of these models restricts their real-world applicability.

Foundation models have emerged as an alternative, pre-trained on massive ECG recordings and applicable to various tasks, including ECG classification and abnormal cardiac troponin detection. One notable example is the ECG-FM model, which is pretrained on more than 2 million ECG segments. The initial evaluation of the model across 13 ECG classes has shown an average accuracy of 94.2%, precision of 80.3%, recall of 88.8%, F1-score of 85.1%, specificity of 95.0%, and AUC-ROC of 97.2%.

Motivated by these results, this study aims to fine-tune ECG-FM on publicly available ECG datasets and compare its performance against conventional ECG classification models. While fine-tuning and final comparisons are still in progress, we hypothesize that the fine-tuned ECG-FM model will achieve superior performance compared to conventional ECG classification models, while reducing development time and computational costs. This study may set the stage for adaptation of foundation models in clinical settings and replace the need for developing task-specific models in the cardiovascular field.