AnyECG: Foundational Models for Electrocardiogram Analysis

1. The authors did not compare AnyECG with contrastive learning-based methods, which are the most related works to AnyECG.

2. It is hard to say that the datasets used in the experiments are representative enough. The datasets used in this work are only a part of the CinC2021 challenge dataset (ref. https://physionet.org/content/challenge-2021/1.0.3/, the Chapman-Shaoxing and Ningbo Database and The University of Michigan (UMich) Database are not used in this work), which are all standard 12-lead ECG datasets. The authors should use more comprehensive datasets to make the name "AnyECG" more convincing, for example, ambulatory ECG datasets collected from wearable devices, or datasets of much larger size (https://zenodo.org/records/4916206), etc.

3. The terminologies used in this work are messy. For example, on page 3, in the paragraph just beneath Figure 1, "w" and "s" are used interchangeably to represent the length of the patch. The authors should use a consistent notation.

1. The authors appear to lack foundational understanding of ECG, leading to a motivation and set of four challenges that do not align with the actual difficulties in ECG tasks. Their data processing and methods rely heavily on conventional NLP approaches, without adequately considering their applicability to ECG tasks. For example, padding 10-second, single- or three-lead ECG data to 12 leads over 60 seconds substantially increases data volume and may lead to erroneous lead-dropout interpretations. Furthermore, the direct conversion of patches from different times and channels into a single sequence risks confusing temporal and channel correlations post-transformer processing, which contradicts realistic ECG analysis logic.

2. The morphology decoder faces similar issues: masking the temporal signal alters the fragment’s attributes, potentially shifting to other ECG abnormalities, thereby hindering accurate reconstruction of temporal signal features.

3. The experimental design does not sufficiently demonstrate the model’s effectiveness in addressing real-world ECG challenges. The results focus excessively on technical validations, lacking a clear analysis of the model’s practical impact on ECG tasks. From the results, it is unclear whether this model qualifies as a foundation model.

1. Lack of Experiments Addressing Key Challenges:

   • The paper does not provide sufficient experiments to verify whether the mentioned challenges are effectively resolved. The ablation studies are inadequate, with none in the main text and only limited ones in the appendix.

2. Insufficient Baseline Comparisons:

   • Despite aiming to position itself as a foundation model, the paper fails to compare its approach with other foundational or self-supervised learning (SSL) methods, which is a critical oversight.

3. Incomplete Downstream Task Descriptions:

   • The explanation of downstream tasks is lacking. It is unclear if the same datasets used for pre-training were also used for downstream tasks. Specific details on the datasets, labels, and tasks (e.g., anomaly detection) are missing.
   • Not all the datasets have the labels for anomlay. What is the downstream dataset for anomaly detection?
   • There is no clear explanation of how corrupted lead generation was conducted or the settings involved (e.g., generating precordial leads from limb leads).
   • Did you use the morphology decoder for the generation of corrupted leads?
   • The methodology and setting for Ultra-Long ECG Recognition are not described, and the meaning of "adaptation" in Table 5 is not explained in the text.
   • Moreover, in Section 7.2, the evaluation metrics for Ultra-Long ECG Recognition are listed as PSNR, SSIM, and MAE, but this does not align with what is shown in Table 5. Which one is correct?

4. Presentation Issues:

   • The notation for patch size is inconsistent, using both $w$ and $s$ interchangeably.
   • Figure 1 (bottom-left) mentions using N transformer blocks, but this contradicts the text, where N refers to the total number of patches.

The authors propose a method to address heterogeneity in ECG data, but the dataset used for downstream applications does not sufficiently reflect this heterogeneity. The model should be tested not only on standard 12-lead ECGs but also on ECGs obtained from diverse measurement devices, including treadmill tests, Holter monitors, patches, and consumer-driven mobile ECG devices.

Using 1-second tokens could be limiting, as electrical signals in ECGs can display features in the range of tens to hundreds of milliseconds. Therefore, it is necessary to verify whether the model can effectively classify critical cardiac abnormalities, such as acute myocardial infarction or acute QT prolongation, which occur on a sub-second timescale in addition to arrhythmias.

According to the paper, the anomaly detection task is described as identifying irregularities in ECGs; however, these irregularities are characteristic of arrhythmias, making it challenging to consider this as a separate task.