DAAC: Discrepancy-Aware Adaptive Contrastive Learning for Medical Time series

Dear Reviewer,

Thank you very much for your thoughtful review and constructive feedback on our submission. We truly value the time and effort you have dedicated to evaluating our work. Please find our point-by-point responses to your comments below.

Q1: Discrepancy signal granularity

A1: Thank you for raising this insightful question. We intentionally design the discrepancy feature as a sequence-level reconstruction error to achieve greater robustness and generalization, especially in the first phase of our three-stage training pipeline. This design aligns with our coarse-to-fine learning strategy, where early stages focus on global guidance and later stages handle finer representations. While finer-grained signals (e.g., point-wise or channel-wise errors) may capture local anomalies, the presence of distribution shifts between external and target datasets can make these fine-grained discrepancies prone to overfitting. This may introduce harmful biases into the downstream representation learning of the ACL module and compromise its generalization ability. To validate our design choice, we implemented and compared the following discrepancy feature variants. Results are included in Appendix E Discrepancy Estimator Study: As shown in the table above, while alternatives Channel-wise MSE vector and Point-wise error sequence capture more local detail, they yield inferior performance compared to our scalar sequence-level discrepancy. We attribute this to their sensitivity to non-generalizable local noise patterns from external data, which can introduce instability in early-stage training and hinder the learning of transferable features. Moreover, we emphasize that this discrepancy signal is only used in the first stage as a weak, auxiliary cue. It is later refined and complemented by the multi-level Adaptive Contrastive Learner (ACL), which captures detailed temporal and spatial structures. Therefore, our choice reflects a principled balance between stability, generality, and progressive refinement, and we find it to be both effective and practical across diverse tasks.

Q2: Loss complexity and its hyperparameter stability

A2: We appreciate this constructive comment and have added a detailed sensitivity analysis in Appendix G.2. To clarify, the 1:1:1:1 weights for the four hierarchical loss components follow prior work [2], which established this as a stable default. Our analysis in added Appendix D2 further supports this configuration's robustness. For the view-level loss $L_V$, we conducted additional experiments by varying its weight. Limited by the size request, we can only put part of the result herein, more results will be shown in added Table D1,The key findings are:

• Effectiveness: Including $L_V$ significantly improves performance, confirming its necessity.
• Saturation: Performance gains diminish as the weight increases, plateauing around a weight of 2.
• Robustness: A weight of 2 provides consistently strong results across all datasets.

In summary, our proposed loss design, including the weight for $L_V$, is both effective and robust. We have noted that a joint optimization of all weights is a promising direction for future work.

Q3: About the risk of distribution mismatch between external data and target data

A3: Thank you for this thoughtful and important question. We agree that relying on external normal data assumes a certain level of domain alignment. In practice, however, our external and target datasets exhibit notable differences, including variations in population demographics and acquisition protocols. For example, in the AD task, the external dataset ([1]) has an average subject age of 63.6 years, while the target dataset ([2]) has an average age of 72.5 years. When training the Discrepancy Estimator (DE) using this external dataset, we indeed observe signs of negative transfer—specifically, in Table 2, the AUROC of COMET+DE drops from 94.44 to 93.97. Despite this, DAAC still outperforms both COMET+ACL and COMET+DE in the same setting, as shown in Table 2. This suggests that the Adaptive Contrastive Learner (ACL) effectively mitigates the potential negative impact introduced by domain-shifted external data, and in some cases, even corrects for suboptimal guidance from the DE. This robustness highlights the strength of DAAC’s multi-stage design in handling real-world domain variability. We have added the related discussion in Appendix C3.

Q4: The proposed method was tested on only 2 medical modalities, EEG and ECG. What are the next target modalities in the medical domain? How about outside of the medical domain?

A4: Thank you for this constructive question. Although our current experiments are limited to EEG and ECG modalities, DAAC is designed as a general and extensible framework for learning representations from hierarchical electrophysiological signals, and is not restricted to specific data types. In future work, we plan to extend DAAC to additional modalities that share similar structural properties, including but not limited to: EMG, ECoG, MEG, fNIRS, PPG, respiratory signals, and GSR. These modalities are widely used in clinical diagnostics, cognitive neuroscience, and human-computer interaction, and typically exhibit clear subject， trial, epoch, and temporal organization, which aligns well with DAAC's multi-level contrastive learning strategy. We believe that applying DAAC to these broader signal types will not only validate the method’s generalizability, but also greatly enhance its practical utility across diverse real-world applications. While our current work focuses on medical time-series, we believe DAAC can also be applied to domains such as fault diagnosis, predictive maintenance, and battery life prediction. These tasks often involve hierarchical time-series and have abundant healthy data but limited labeled faults—conditions that align well with DAAC’s design, particularly the use of external normal data for discrepancy estimation. We see this as a promising future direction.

Q5: The authors state that “We explicitly discuss limitations in Section 6”, but the main text ends with “5 Conclusion”. Where is the section 6?

A5: Thank you for pointing this out. This was a formatting oversight in our submission. The limitations are indeed discussed in the final part of Section 5 (Conclusion). We have corrected the section reference and will ensure that this is clearly marked in the revised version.

Your insights have been instrumental in helping us improve the clarity and rigor of our paper. We have carefully addressed each concern and believe that the revised version better reflects the contributions and robustness of our approach.

Sincerely,

The Authors

References [1] Miltiadous, Andreas, Katerina D. Tzimourta, Theodora Afrantou, Panagiotis Ioannidis, Nikolaos Grigoriadis, Dimitrios G. Tsalikakis, Pantelis Angelidis et al. "A dataset of EEG recordings from: Alzheimer’s disease, Frontotemporal dementia and Healthy subjects." OpenNeuro 1 (2023): 88. [2] Escudero, J., Abásolo, D., Hornero, R., Espino, P. and López, M., 2006. Analysis of electroencephalograms in Alzheimer's disease patients with multiscale entropy. Physiological measurement, 27(11), p.1091.

Dear Reviewer,

We extend our sincere appreciation for your time and expertise in reviewing our work. Our detailed responses to the reviewers' comments are presented below.

Q1: Minor issues – typo in ‘surpervised’; unclear bold formatting in Table 2.

A1: Thank you for pointing these out. We have corrected the typo "surpervised" to "supervised" in the revised figure. Also, We have clarified the table formatting: "boldface now consistently indicates the best performance per task. underline denotes the second best". Meanwhile, we have updated the captions of Table 1 and Table 2 accordingly to avoid term confusion, which will largerly improve our clarity. We appreciate your careful reading and helpful suggestions.

Q2: Please define the absolute size (number of subjects / images) and class balance of the external normal pool used in each experiment. In addition, supply an ablation (e.g., 10 %, 25 %, 50 %, 100 % of your current pool) showing how downstream performance varies with normal-set size.

A2: We appreciate the reviewer’s concern regarding the feasibility and value of using a large-scale healthy population pool. To address this, we conducted ablation experiments on all three datasets (AD, PTB, TDBrain), varying the proportion of healthy individuals used for pretraining (e.g., 5%, 25%, 50%, and 100%), as shown in Table E3. For each proportion, we further evaluated performance under different amounts of labeled training data (5%, 25%, 50%, 100%). Due to space constraints, we only present the results for the AD, PTB dataset in the main text, but consistent trends were observed across all datasets. We have updated Appendix E3 to clarify this point and provide additional results. Key findings:

• Across all datasets and settings, increasing the size of the healthy population consistently leads to better model performance.
• The improvement is existed even in low-resource settings, where the contrastive pretraining helps compensate for limited supervision.
• The performance trend is stable and monotonic, suggesting that more external healthy data continues to benefit the model without introducing instability.
• These results confirm that the proposed approach can adapt to different scales of external data and highlight the practical value of incorporating healthy population information, even when collected in limited amounts. This validates the scalability and deployment feasibility of our method, and demonstrates that the use of external healthy individuals is both effective and justifiable.

Q3: Provide quantitative evidence (e.g., AUROC, distribution plots) that healthy samples indeed yield lower reconstruction error E than abnormal samples across all datasets.

A3: We appreciate the reviewer’s suggestion and have conducted additional analysis to empirically validate the assumption. Due to the submission policy disallowing figures in the rebuttal, we provide a textual description of the key results and refer the reviewer to Figure F1 (has been added to the revised manuscript) in Appendix F.2 for visual confirmation.

We performed Kernel Density Estimation (KDE) on the reconstruction errors of both normal and abnormal samples, and showed in Figure F1. The normal samples exhibit a sharp peak centered around a low reconstruction error (~0.2), with very narrow spread, indicating consistent low-error reconstructions. In contrast, the abnormal samples follow a positively skewed distribution, with reconstruction errors more broadly distributed and a long tail extending toward higher values (~1.0). This empirical result confirms that the model consistently assigns higher reconstruction errors to abnormal samples.

Furthermore, using the reconstruction error as the anomaly score yields an AUROC of 0.974 on the AD dataset, quantitatively supporting the model’s discriminative capability between normal and abnormal samples.

Q4：Supply a concise table (layers, filter sizes, activation functions, parameter count) for both encoder and decoder, plus a short justification for key design choices (e.g., depth, latent dimension).

A4: We appreciate the reviewer’s feedback regarding the architectural transparency of our model. To address this, we have included detailed architectural specifications in Table G1 now clearly outlines the encoder type, dimensional configurations (input/output/hidden/channel/head), activation function, and parameter count. These additions ensure that all design choices—layer counts, attention structure, and parameter sizes—are explicitly stated.

Q5: The privacy considerations involved even when using “healthy” clinical data.

A5: We agree that even healthy clinical data may raise privacy concerns. As noted in our experimental section, all datasets used in this study are publicly available and had received institutional review board (IRB) approval prior to release, which ensures appropriate de-identification and regulatory compliance (e.g., HIPAA, GDPR).

Reply to the concerns on significance and clarity

We really appreciate for your valuable insights. These comments allow us to further improve our work quality via revision. We hope the responses have fully addressed your questions and provide strong support for the significance of our method. With all these designs, our approach demonstrates significant performance improvements across multiple datasets (AD, PTB, TDBrain) even in a low-resource setting. For example, on the AD dataset, even with only 5% of external healthy data, the model could significantly outperform the baseline without DE pretraining. This validates the effectiveness and practical adaptability of our framework. In addition, we really care about your feedback regarding clarity. Accordingly, we have revised our paper seriously, which includes clarifying the method definitions (see Table 1), architectural details (see Table G1), and experimental setup (see Appendices E and F), etc. In addition, a native speaker is asked to further improve the narratives in the manuscript. We suppose these endeavors could help to alleviate your clarity concern. Once more, thank you very much for your time and suggestions for reviewing our paper.

Best regards,

The Authors

Dear Reviewer,

We sincerely appreciate your thoughtful review and helpful suggestions. We have addressed all comments and provided detailed responses below.

Q1: This paper use many terms refers data types (e.g., normal/external/target/original data) and includes unexplained abbreviations (especially in ablation studies), leading to confusion about the Discrepancy Estimator’s functionality and experimental setup.

A1: Thank you for your suggestion. We have revised the model names for consistency and clarity, and provided clear definitions for each model in the captions of Table 1 and Table 2. Also, we additionally provide explanations for all key terms about data types in Appendix C. In the medical field, datasets from different hospitals often exhibit a certain degree of distributional shift. Models trained on data from one hospital may demonstrate unstable performance when applied to others. To enhance model generalization, we propose leveraging DE (Discrepancy Estimator) to incorporate data from other hospitals. Although distributional differences exist, we argue that the sequential patterns of normal data (i.e., data from healthy patients) can still provide valuable guidance for diagnosis in the target hospital. In this study:

• The dataset from the current hospital is referred to as the internal dataset (or target dataset), which participates in the second and third stages of model training.
• Datasets from other hospitals, used for guidance, are termed external datasets which are from healthy patients and contribute to the first stage of training.

Q2: The discrepancy is just an augmented feature of the original data, so why do you choose discrepancy as this additional fearture? what about other metrics like the data point to the cluster of normal/disease data?

A2: Thank you for raising this thoughtful question. We agree that the discrepancy feature serves as an auxiliary signal to enrich the original data representation. Our choice to use the sequence-level reconstruction error stems from the need for a robust signal that generalizes well under the potential distribution shift from the external dataset. Specifically, this design:

• Avoids overfitting to local noise or domain-specific patterns, which can occur when using fine-grained or structure-sensitive alternatives (e.g., point-wise or cluster-based distances).
• Aligns with our coarse-to-fine training strategy, where this global signal (i.e., discrepancy) guides the early stage and is progressively refined by the Adaptive Contrastive Learner (ACL) in later stages. To validate this choice, we further conducted ablation experiments (see Appendix F) comparing our method with alternative discrepancy features, including:
• Channel-wise MSE vector (16-dim),
• Point-wise reconstruction error sequence (256-dim),
• Cluster-wise latent distance to normal data. As shown in Table F1, while these alternatives incorporate more structural or local cues, they consistently underperform compared to our scalar discrepancy. We attribute this to their higher sensitivity to noise and domain shift, which can degrade early-stage learning. In summary, our discrepancy design is not arbitrary, which reflects a careful tradeoff between stability, informativeness, and generalizability, and is empirically validated across diverse datasets and settings.

Q3: Which part do you think contribute the most to the final improvement?

A3: Thank you for the valuable question. As shown in Table 2, the Adaptive Contrastive Learner (ACL) provides the most consistent and significant improvements across all datasets under the full supervision setting (100% labeled data). However, under limited-label conditions (e.g., 10% labeled data), the Discrepancy Estimator (DE) becomes increasingly important. Therefore, the performance gain of DAAC comes from the synergistic integration of both components, rather than relying on a single dominant module. Specifically, DAAC’s effectiveness is amplified through:

• Coarse anomaly guidance from DE, which helps ACL focus on more informative signals.
• Multi-head adaptive view construction in ACL, enabling more diverse and robust contrastive pairs for better representation learning. We emphasize that the largest performance gains emerge when DE and ACL are progressively combined in a coarse-to-fine manner, as demonstrated by the full DAAC outperforming partial variants such as COMET+DE and COMET+ACL.

We are deeply grateful for the your constructive feedback. We hope that our detailed response provided above can address your questions and suggestions.

Best regards,

The Authors

Dear Reviewer,

Thank you for your thoughtful and detailed comments. We have carefully addressed each of your concerns below and revised the manuscript accordingly.

Q1: Concern regarding whether more established datasets, such as MIMIC, were considered for evaluation.

A1: Thank you for this question. While MIMIC is a valuable dataset for EHR studies [1], our model, DAAC, is specifically designed for multi-level electrophysiological time-series (e.g., EEG, ECG) with inherent hierarchical structures (subject, trial, epoch, temporal). The selected AD, PTB, and TDBrain datasets exhibit these structures, providing a suitable testbed for DAAC. Their complexity and use in established benchmarks like COMET [2] ensure a proper and sufficient evaluation of our method's effectiveness (Appendix A.1).

Q2: Regarding the sensitivity of the contrastive loss weights.

A2: We appreciate this constructive comment and have added a detailed sensitivity analysis in Appendix G.2. To clarify, the 1:1:1:1 weights for the four hierarchical loss components follow prior work [2], which established this as a stable default. Our analysis in added Appendix D2 further supports this configuration's robustness. For the view-level loss $L_V$, we conducted additional experiments by varying its weight. Limited by the size request, we can only put part of the result herein, more results will be shown in added Table D1,The key findings are:

• Effectiveness: Including $L_V$ significantly improves performance, confirming its necessity.
• Saturation: Performance gains diminish as the weight increases, plateauing around a weight of 2.
• Robustness: A weight of 2 provides consistently strong results across all datasets.

In summary, our proposed loss design, including the weight for $L_V$, is both effective and robust. We have noted that a joint optimization of all weights is a promising direction for future work, and has been mentioned in revision.

Q3: Details of training process.

A3: Additional training details are now in Appendix G.1. All experiments were run on an NVIDIA RTX 4090 GPU. Our training pipeline is efficient, and the final model is lightweight for deployment. We will release the full code upon acceptance.

Q4: Using a single sequence-level reconstruction error as the discrepancy feature might be too coarse.

A4: We intentionally use a sequence-level error to ensure robustness during the initial training phase, as this coarse-grained signal prevents overfitting to noise or center-specific patterns from external data. While finer-grained measures like point-wise errors can be sensitive to physiological fluctuations and sensor artifacts [3], a sequence-level discrepancy captures global deviations more reliably. Our experiments (Appendix F) confirm that alternatives like channel-wise or point-wise errors yield inferior performance due to their sensitivity to local noise. This coarse feature acts as an auxiliary signal, and we designed the downstream Adaptive Contrastive Learner to subsequently extract fine-grained representations across multiple levels, ensuring detailed information is fully utilized.

Q5: Model names in the results tables are unclear.

A5: Thank you for the suggestion. We have revised the model names for consistency and provided clear definitions in the captions of Table 1 and Table 2.

Q6: The Discrepancy Estimator relies on large external normal datasets, which may be infeasible for rare diseases.

A6: This is an important point. (1) For the common diseases in our study (Alzheimer’s, Parkinson’s, etc.), collecting normal data is generally feasible and low-cost. (2) For rare diseases, we agree acquiring large-scale data is difficult. To assess this, we conducted an ablation study and found that even with only 5% of the healthy data, our model still achieves performance gains. This suggests the Discrepancy Estimator remains beneficial even in resource-constrained scenarios.

Q7: Could you discuss the potential risk of negative transfer from domain shift?

A7: We agree this is a key consideration. Our external and target datasets do exhibit domain shifts (e.g., average subject age of 63.6 vs. 72.5 years in the AD task [4, 5]). We observed signs of this negative transfer when applying a baseline model, where the AUROC dropped from 94.44 to 93.97 (Table 2). Despite this, DAAC still outperformed all baselines in the same setting. This demonstrates that our Adaptive Contrastive Learner (ACL) effectively mitigates the negative impact from domain-shifted data, showcasing the robustness of DAAC's multi-stage design. We have added this discussion to Appendix C3.

Q8: The novelty seems limited as the work builds heavily upon COMET.

A8: We acknowledge that we build upon COMET's framework. However, the core innovation of DAAC is the synergistic integration of the Discrepancy Estimator (DE) and the Adaptive Contrastive Learner (ACL) into a three-phase, curriculum-inspired framework. This design follows a coarse-to-fine progression:

• Phase 1: DE learns a coarse, abnormality-aware signal from external data.
• Phase 2: ACL uses this signal to dynamically discover and contrast fine-grained views.
• Phase 3: The encoder is fine-tuned for the downstream task. The synergy of these components is vital. Our ablation studies (Table 2) show that simply adding DE or ACL to COMET yields accuracies of 84.82% and 91.16% respectively, while the integrated DAAC framework achieves 93.23%. This demonstrates that the framework itself is a meaningful advancement.

Thank you again for your patience and invaluable feedback.

Best regards,

The Authors

References

[1] Johnson, A. E. W., et al. (2021). MIMIC-IV, a freely accessible electronic health record dataset. Scientific Data.

[2] Wang, Y., et al. (2024). Contrast everything: A hierarchical contrastive framework for medical time-series. NeurIPS.

[3] Liu, Z., et al. (2023). Self-supervised contrastive learning for medical time series: A systematic review. Sensors.

[4] Miltiadous, A., et al. (2023). A dataset of EEG recordings from: Alzheimer’s disease, Frontotemporal dementia and Healthy subjects. OpenNeuro.

[5] Escudero, J., et al. (2006). Analysis of electroencephalograms in Alzheimer's disease patients with multiscale entropy. Physiological Measurement.