Interpretable Pre-Trained Transformers for Heart Time-Series Data

We thank the reviewer for highlighting the interpretability of our network as a strength, and specifically down to the level of decoding of individual attention heads that respond to physiologically relevant features in PPG and ECG. We have responded to the reviewer’s comments below and believe that they have improved the quality of our manuscript.

• The explainability method of GPT models was shown to focus on previous cycles meaning that it can observe a beat w.r.t the previous one which is why the downstream task interprets well where distance between two consecutive beat is important criteria. However, this might not be the case for other common tasks such as sleep staging where the input in 30 second ECG or PPG signal and separating sleep stages such as wake, light sleep, deep sleep and REM sleep manifests from HR variability. Another experimentation for a downstream task would be interesting to see if the proposed explainability idea fits well to provide clinical domain specific interpretation, given the fact that GPT models were found to be useful in these downstream tasks.

Thank you for this suggestion. In our experience with HRV, we have found that generally inputs require time greater than 2 minutes, and usually 5 minutes, to account for the low frequency variations between 0.04 Hz (VLF band in HRV) and 0.1 Hz (LF band in HRV). Even at 30 seconds, this would require a 3-fold increase in context length for our current PPG model, and a 6-fold increase for our current ECG model. To maintain efficiency, implementing such a task would likely require the use of algorithms such as sparse attention (Roy et al, 2021) [1]. We have added this to a future work section, in Appendix A.13.

For a clinical domain specific task, and to demonstrate that our model’s interpretable attention mechanism can focus on morphological features as well as timing-based features, we have added an extra downstream task for the classification of premature ventricular contractions. Premature ventricular contractions manifest in clearly visible morphological changes in the ECG waveform, and thus represent a clear litmus test for the interpretability of our model. As with all other tasks, we test only on unseen subjects. Our interpretability analysis (see Figure 10, Appendix A.10) reveals that, as desired, attention clearly shifts to regions where PVCs are present.

• Few related GPT interpretability studies were referenced. More references should be included to have a broad picture based on GPT interpretablity. For example, a. Creswell, A., Shanahan, M., & Higgins, I. (2022). Selection-inference: Exploiting large language models for interpretable logical reasoning. arXiv preprint arXiv:2205.09712. b. Ben Melech Stan, G., Aflalo, E., Rohekar, R. Y., Bhiwandiwalla, A., Tseng, S. Y., Olson, M. L., ... & Lal, V. (2024). LVLM-Intrepret: an interpretability tool for large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 8182-8187).

We have expanded our introduction (additions highlighted in blue), and added Appendix A.14 to clarify our attention method relative to these references. We have made sure to specifically mention LVLM-interpret, which exhibits interpretability in images with a language prior, as this shares parallels with the interpretation of physiological time series with a “classification” prior.

• Figure 3 captions needs to be corrected a-d.

Thank you, we have corrected this figure caption.

We appreciate the reviewers feedback and positive comments about the strengths of our work, including agreement with the clear interpretability of our network. We also appreciate the reviewer’s suggestions for improvements and believe that they have improved the quality of our manuscript.

1. The datasets used in the experiments are not comprehensive enough. The authors used the CinC2020 dataset, which is superseded by the CinC2021 dataset. The latter dataset is more comprehensive and contains more data. Moreover, there are other larger datasets available, such as the CODE-15% dataset (https://zenodo.org/records/4916206), etc.

Thank you for this suggestion. In our pre-trained model, we purposefully used a similar ratio between the number of training tokens and model size to what is currently used in large language model training, such as in Llama. We have referenced this in Subsection 3.2 “Architecture and training”. For pre-training, adding more data than was available in CinC 2020 would be necessary only when scaling the model size. However, we find that for most tasks this may not be necessary as we achieve our goals of i) interpretability and ii) comparable and even SOTA performance in all downstream tasks. For fine-tuning, we recognise the importance of evaluating our model on this more comprehensive dataset to enhance result reliability. Thus, we have added two new fine-tuning tasks that utilise data that was added in the expanded CinC 2021 dataset, namely the Chapman 12-lead ECG dataset.

2. The authors did not compare their method in their numerical experiments with other representation learning methods, such as contrastive learning-based methods (for example CLOCS (https://proceedings.mlr.press/v139/kiyasseh21a/kiyasseh21a.pdf)), to show the effectiveness of their method.

3. Not enough downstream tasks are conducted to evaluate the learned representations (or the pre-trained GPTs).

We thank the reviewer for this suggestion. We have implemented two more fine-tuning tasks for ECG, taking our total number of downstream tasks to five. We have specifically added the 4-class arrythmia classification with the Chapman dataset to directly compare our performance numerically to CLOCS, which beats other contrastive learning-based methods. We have included our performance on this task and other downstream tasks in a new summary table, in Appendix A.11, as well as a summary paragraph of comparisons to other architectures in the literature. We have also expanded our related work section to include works such as Kiyasseh et al 2021.

Whilst the performance of our model is comparable with the state-of-the-art performance in methods suggested by the reviewer for all downstream tasks, the paramount point is that no works, in either our previous literature review or our now extended literature review, have convincing interpretability. This remains the key contribution of our work. To this end, we have also included further interpretability analysis for the downstream task of detecting premature ventricular contractions (PVCs), using a subset of the Chapman dataset (see Appendix A.10, and Figure 10). Whilst the interpretable attention of our AF downstream task focused exclusively on timing, the attention in the PVC task highlights the distinct changes in morphology that occur with a PVC. This observation further highlights the robustness and adeptness of our model's interpretability in handling diverse physiological signal characteristics, owing to the training method we employed.

In our final submission, we will provide a public repository, where we open source all the model files and code, and even provide a suite of GUIs designed for interpretability. This is aimed at making our models easy to implement, for both AI researchers and medical professionals, and in this way is an attempt to bridge the gap between AI research in healthcare and real-world implementation.

We appreciate the reviewer’s feedback, including highlighting the interpretability of our work as a strength. We have responded to all of the reviewer’s comments below, and believe that the reviewers comments have positively impacted our manuscript.

1. The model's novelty is questionable, as the PPG-GPT has previously been explored across multiple tasks (Chen et al., 2024). It will be helpful to specify how the current work differs from it.

We thank the reviewer for pointing us to this work by Chen et al. The major difference between our work and the previous PPG-GPT work is that they tokenise based on patches, as is common in time series transformers, rather than based on individual time points to form a finite vocabulary, as is the case with our work. Moreover, rather than predicting the next token like GPT, they interpolate patches bidirectionally in a similar way to BERT. Importantly, in the case of Chen et al, this leads to a network which does not offer interpretability. We anticipate this is due to i) a significantly larger model implemented combined with patch processing and ii) to the loss function used in training. Interpretability is critical for the acceptance of AI models in healthcare, and this is why we have made these specific design choices in our work in order to create interpretable models. We have updated our related work section to include this discussion (highlighted in blue).

2. The manuscript does not compare models on different PPG or ECG tasks compared to previous PPG-GPT (Chen et al., 2024). The comparison with one baseline per PPG and ECG is mentioned only in the appendix. Importantly, the models are compared on performance using different setups.

We thank the reviewer for this suggestion. We have now extended our downstream tasks to further evaluate our models and provide more comparisons in terms of accuracy with other pre-trained models in the literature. We have focused on implementing more arrythmia based tasks, namely 1) 4-class arrythmia detection (supraventricular tachycardia, atrial fibrillation, sinus bradycardia and sinus irregularity), 2) detection of premature ventricular contractions. These tasks were chosen due to their medical relevance and suitability for an interpretable model capable of aiding medical practitioners. On the 4-class arrythmia detection in ECG, we achieve an AUC of 0.97 with just a single lead, that beats popular contrastive learning-based methods (e.g CLOCS, Kiyasseh et al 2021 [1], AUC 0.90). On beat detection in finger PPG, our average F1 score of 0.98 corresponds to state-of-the-art performance compared with MSPTD (Bishop et al, 2018 [2]) . On the PVC task, we extend our interpretability analysis and achieve an AUC of 0.99 (Appendix A.9, A.10, A.11).

3. The experiments have not been performed across multiple runs, so the variability (std or IQR) of the performance is not clear. Consider 5-fold or 5 seeds.

4. Consider additional metrics, for example, sensitivity, specificity, false-positive rate, false-negative rate, and F1 score for AF.

Thank you for these suggestions, we have now performed all fine-tuning tasks with 5 random seeds, to produce means and standard deviations for each metric. For all relevant tasks, we have also included all additional metrics suggested. These are summarised in a table in Appendix A.11.

5. The majority of the paper focuses on interpretability. However, it is difficult for the machine-learning community to validate clinical utility claims that require experience in reading ECG or PPG. Have you collaborated with clinical experts to validate the interpretability claims? Could you provide more evidence that the examples you have shown guarantee the presence of AF?

To demonstrate interpretability, we purposefully chose AF, given that it is a very straightforward arrythmia to see visually. To ensure that this is communicated appropriately to machine learning researchers without medical training, we have included a thorough description of AF with citation (Wijesurendra and Casadei, 2019) [3] in Section 6, characterised as an irregular heart rate which manifests itself in rapid increases and decreases in heart rate. Our attention analysis shows that the model shifts attention specifically to regions where heart rate rapidly decreases or increases, and we have annotated this on Figure 5. Further to this, we have also extended our interpretability analysis to premature ventricular beats, which manifest clearly through morphological changes in the ECG. The attention of our model in this case shifts directly to regions with premature ventricular contractions. This extra analysis is present in Figure 10 (Appendix A.10). Note that in Figure 10 on the left there are two clear PVCs, and the fine-tuned attention mechanism spikes directly for both. On the right, there is one example of a PVC, and the attention shifts to just this one PVC.