acceptodds
Under review as a conference paper at ICLR 2027

HB-IGCL: Heart-Brain Instance-Global Contrastive Learning for Emotion Recognition using Long-duration ECG and EEG

Abstract

Emotion recognition is a key topic in affective computing, where electrocardiogram (ECG) and electroencephalogram (EEG) provide complementary peripheral and central physiological views. For long-duration recordings, a common practice is to divide each recording into short segments and train segment representations with the recording-level emotion label. This practice is convenient but conceptually restrictive: it assumes instance-level emotion supervision that is not actually observed, while emotional physiological responses may vary substantially across time. In addition, many ECG-EEG methods process the two modalities independently or combine them only through late fusion, underusing the natural correspondence between synchronized cardiac and neural observations. To address these issues, we propose Heart-Brain Instance-Global Contrastive Learning (HB-IGCL). Rather than treating every segment as an emotion-labeled sample, HB-IGCL learns whether a local physiological instance is compatible with the paired global context from the other modality within the same synchronized recording. ECG-to-EEG and EEG-to-ECG objectives provide bidirectional local-to-global contextual alignment, and a spatial-guided alignment module (SGAM) further relates lead-level spatial structure to global temporal context. We evaluate HB-IGCL in subject-independent experiments on AMIGOS, DREAMER, and MPED. Results show that the proposed framework improves representation learning across multiple encoders and datasets.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Emotion recognition, ECG, EEG, Contrastive learning, Representation learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_CQ9DOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes HB-IGCL, a self-supervised pretraining framework for emotion recognition from long ECG–EEG recordings. Instead of assigning the recording-level label to every short segment, each recording is split into non-overlapping instances and treated as a bag. Two contrastive objectives align local instances of one modality with the global representation of the other modality from the same recording, using other recordings in the batch as negatives. C2E contrasts ECG instances with the averaged EEG global representation. E2C contrasts EEG instances with an ECG global representation weighted by the C2E scores. A spatial-guided alignment module (SGAM) additionally aligns a lead-level spatial representation with the global temporal one. The framework uses four encoders (two per modality), a proposed multi-branch instance encoder (AMIE), and a gate-and-refinement (GR) module. For downstream classification, the pretrained encoders are frozen and combined through a cross-attention head and a classifier. Experiments on AMIGOS, DREAMER (4-class), and MPED (3-class) under subject-independent splits compare HB-IGCL with three pretraining frameworks (LMM, EMM, CCM) across five encoders. Among these frameworks, HB-IGCL gives the best F1 for every encoder.

Critical strengths
  • The motivation is reasonable. Assigning the recording-level label to every short segment is a real issue for long physiological recordings, and using the within-recording ECG–EEG correspondence as a training signal is a sensible alternative.

  • The experimental effort is substantial: five encoders, three datasets, five seeds, and subject-independent evaluation. The authors also adapt the baselines to their setting and report the full sweeps (App. F–H) rather than only the selected configurations, which makes the comparison more transparent.

Critical weaknesses
  1. The gain from pretraining is inconsistent across datasets. Compared with training the same model from scratch (Rand. Init., Table 2), frozen HB-IGCL features improve F1 clearly on AMIGOS (+9.9) and modestly on DREAMER (+3.8), but are slightly worse on MPED (84.66 vs. 85.86). With full fine-tuning, the differences are +5.6, −0.5, and +3.7, mostly within the reported standard deviations. Pretraining does reduce variance in several settings, but part of the performance seems to come from the encoder and downstream architecture. A supervised-from-scratch row for each encoder in Table 1 would make the contribution of pretraining clearer.

  2. Role of the fusion head. In Table 3, the single-modality variants are close to chance, and concatenating both modalities gives similar results (27.45 F1 on AMIGOS, 42.24 on MPED), while the cross-attention head reaches 66.23 and 84.66. Y_c and Y_e are described as representation vectors (L269–273). If cross-attention is computed over a single key token, it reduces to a projection of the EEG features, so it is unclear what produces such a large gap. Clarifying the head design, its parameter count, and whether the single-modality variants use the same frozen encoders would help interpret how much of the result reflects the pretrained representations.

  3. Inconsistent results across tables. The “GR” rows of Table 4 (DREAMER) report 42.47 / 60.55 (AMIE), 24.80 / 46.39 (ModernTCN) and 15.92 / 32.22 (PatchTST). Table 1 reports 50.53 / 63.61, 31.73 / 50.00, and 28.99 / 44.44 for what appears to be the same configuration. I may be missing a difference in setup, but none is described. Within Table 4, “w/o GR” also matches or beats “GR” in F1 for ModernTCN and OS-CNN, so the evidence for GR seems mixed.

  4. Model selection. App. C and H state that α = 0.5, β = 0.8, and the CCM coefficient a = 0.3 were chosen because they give the best results in the reported sweeps (L732–741, L884–885). These values are identical, including standard deviations, to the corresponding test results in Table 1 (e.g., Table 10 vs. CCM-AMIE: 56.07±2.99 / 57.69±2.81). T = 4 s is likewise chosen from Fig. 9. Since a validation split exists, please clarify whether these selections were made on the validation or the test subjects.

  5. Limited evaluation protocol. Each dataset uses a single fixed subject split with few test subjects (5 / 4 / 3; Table 5), and the five seeds only vary the initialization, so the standard deviations do not reflect variation across subjects. K-fold cross-subject or leave-one-subject-out evaluation would make the results more reliable.

  6. Minor

  • The instance-removal analysis (Fig. 6) is on AMIGOS only, without error bars, and is non-monotonic.
Questions for authors
  1. Were α, β, T, and the baseline hyperparameters selected on the validation or the test subjects?

  2. What setup produces the Table 4 “GR” results, and why do they differ from Table 1?

  3. How does a supervised-from-scratch model perform with each encoder in Table 1?

  4. How is the cross-attention head defined? Are Y_c and Y_e single tokens, and how many parameters does it have compared with the concatenation head?

1 comment

Response to Reviewer CQ9D

We thank the reviewer for the thoughtful feedback. New results will be added to the revised manuscript.

1. W1 + Q3 (Supervised-from-scratch controls)

We have added supervised-from-scratch controls for every encoder (Table 8, Appendix E of the revised manuscript); the table reports supervised-from-scratch test F1 and Acc, to be compared with Table 1.

EncoderAMIGOS F1(%)AMIGOS Acc(%)DREAMER F1(%)DREAMER Acc(%)MPED F1(%)MPED Acc(%)
ModernTCN20.36±3.8328.98±3.8029.55±7.8151.67±4.8530.38±5.8851.66±10.80
PatchTST21.72±5.0831.28±7.0021.55±6.8539.72±6.0233.21±4.9356.90±5.79
PatchTST+25.97±3.6734.62±2.0325.08±4.8440.56±6.6135.19±3.0958.33±3.04
OS-CNN21.01±6.8025.38±4.5721.79±5.1541.39±6.0129.32±7.6149.05±12.30
AMIE56.31±6.1160.25±3.3446.75±2.0362.50±0.8885.86±6.0987.62±5.25

2. W2 + Q4 (Definition of the fusion head)

Neither Yi,cY_{i,c} nor Yi,eY_{i,e} is a single token. An ECG signal is first split into instances, each fed into two parallel branches (ECG Encoder + GR module + GAP) that each output a 1×128 tensor; concatenating the two branch outputs yields the final ECG comprehensive representation Yi,cY_{i,c} ∈R2×128ℝ^{2×128}. The EEG signal is processed analogously to obtain Yi,eY_{i,e} ∈R2×128ℝ^{2×128}. We have clarified this in the revised manuscript.

The “concatenation + classifier” head has 2,052 trainable parameters, and the cross-attention head has 199,300 only 2.62% of the model’s ~7.6M trainable parameters), while the performance gain is substantial: F1 rises from 27.45% to 66.23% on AMIGOS and from 42.24% to 84.66% on MPED. All unimodal and multimodal variants in Table 3 use the same pretrained and frozen encoders and the same data split, differing only in fusion strategy; the Table 3 caption has been revised accordingly.

3. W3 + Q2 (The setting behind the GR results in Table 4)

The discrepancy between Tables 4 and 1 stems from the protocol. The original Table 4 reports results with ECG and EEG concatenated before the encoder, without pretraining and with the concatenated input passed directly to the encoder, serving as a quick check of the GR design; Table 1 reports pretraining followed by fine-tuning. The table below reports the GR ablation with pretraining and fine-tuning and replaces the original Table 4 in the revised manuscript.

VariantAMIE F1(%)AMIE Acc(%)ModernTCN F1(%)ModernTCN Acc(%)PatchTST F1(%)PatchTST Acc(%)PatchTST+ F1(%)PatchTST+ Acc(%)OS-CNN F1(%)OS-CNN Acc(%)
w/o GR30.21±5.9846.39±4.2425.71±3.9844.45±6.7324.97±1.6839.45±7.8326.01±5.7339.45±2.5328.90±6.2546.11±4.21
w/o Refine33.36±5.1150.00±6.1331.41±2.5348.33±3.7626.67±6.4243.89±6.2624.62±4.5341.67±2.1929.44±7.0543.33±3.85
GR→TE34.05±3.2645.83±3.1031.82±4.9448.89±1.1628.67±4.3143.61±4.8726.79±3.8642.22±4.7732.21±6.9248.06±7.32
GR50.53±2.8963.61±2.0431.73±5.0550.00±7.0328.99±2.0244.44±3.2926.99±5.0840.83±8.2731.34±1.4548.89±2.55

4. W4 + Q1 (Model selection: validation or test set)

The hyperparameters (α,β,T) were selected on the validation set, and they also yield the best overall performance on the test set (Figures 8 and 9, original manuscript). The corresponding validation-set sweeps, which are consistent with the test-set optima, have been added to Appendices C and D of the revised manuscript as Tables 6 and 7.

The hyperparameters of the baseline models correspond to the best results on the test set; the full results are reported in Tables 7–10 of the original appendix (Tables 10–13 in the revised manuscript). Baseline hyperparameters are set to yield their best test-set results, making the evaluation of HB-IGCL more stringent.

5. W5 (Single fixed split and few test subjects)

Since the three datasets contain only 40, 23, and 23 subjects and all suffer from class imbalance (severe in AMIGOS and DREAMER), random subject-independent splitting may leave certain classes absent from the validation or test subsets. We therefore follow [1] and [2] in assigning subjects to mutually exclusive training, validation, and test subsets. We have added these details to Appendix A of the revised manuscript.

Reference:

[1] Xu et al., IEEE BIBM 2025.

[2] Zhong et al., IEEE BIBM 2025.

Reviewer_hMxZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the limits on emotion recognition using long-duration physiological signals. They proposed a biologically intuitive Bidirectional Instance-Global Contrastive Learning (C2E & E2C) framework. The model learns to match a local instance (e.g., 4s of ECG) with the global context of the other modality (e.g., the entire EEG recording). Unlike models that only focus on time, SGAM uses a Transformer to model the spatial relationships between different leads/channels (heart-brain axis) and aligns this spatial structure with the temporal global context.

The framework was validated on three major public datasets: AMIGOS, DREAMER, and MPED, covering various recording lengths and emotion protocols. The Backbone was tested against 5 different encoders (including ModernTCN and PatchTST). HB-IGCL consistently improved performance regardless of the underlying architecture. Furthermore, they ran ablation studies, suggesting that bidirectional alignment (C2E + E2C) outperforms unidirectional approaches and that the SGAM module provides significant complementary spatial information.

Critical strengths
  1. The paper addresses a meaningful problem: learning from long ECG–EEG recordings without assigning recording-level emotion labels to every segment. Bidirectional instance–global alignment is a reasonable approach.

  2. Experiments cover three datasets and multiple encoders, with additional analyses of training objectives, fusion strategies, and instance duration. The results demonstrate promising downstream utility.

Critical weaknesses
  1. The comparisons mainly cover adapted pretraining frameworks and generic time-series encoders. State-of-the-art methods for cross-subject multimodal physiological emotion recognition are missing.

  2. The evaluation uses a single fixed subject split, with only 3–5 test subjects per dataset. Repeating training with five seeds does not capture variability across subject partitions.

  3. The criticism of masked reconstruction in the introduction is insufficiently substantiated. Mask-ratio sensitivity alone does not establish that performance degradation results from removing sparse emotional evidence. The cited Sarkar and Etemad paper uses transformation recognition rather than masked reconstruction.

  4. I suppose that Table 3 writes "Full" as "All" in the last two lines. The full GR configuration in Table 4 reports different DREAMER results from the main tables, without explaining the discrepancy.

  5. Figure 2 is difficult to follow, particularly the placement and connections of SGAM. It does not clearly show SGAM’s lead-level inputs or distinguish its auxiliary alignment objective from the main representation pathways. A clearer separation of data flow and loss connections would improve readability.

Questions for authors
  1. Can representative domain-specific baselines be evaluated under the same protocol?

  2. Please clarify whether pretraining uses only training subjects, provide the subject IDs and AMIGOS recording-selection criteria, and evaluate robustness across additional subject splits.

  3. What explains the discrepancy between the GR+AMIE results in Table 4 (42.47% F1) and the main DREAMER results (50.53% F1)?

  4. Please clarify SGAM’s inputs and role in Figure 2, and distinguish the pretraining-only components from those retained for downstream evaluation.

4 comments

Response to Reviewer hMxZ

We thank the reviewer for the detailed comments on the experimental protocol and the presentation. We address each point below; the new results are included in the revised manuscript.

1. W1 + Q1 (Domain-specific baselines)

We agree that comparison with domain-specific baselines is necessary. Table 7 (Table 10 in the revised manuscript) reports results for the Criss-Cross Transformer [1]. However, existing baselines in this field, including physiological-signal encoders such as BIOT and Medformer and the ECG-EEG emotion-recognition model Criss-Cross Transformer, all segment ECG and EEG into fixed-length short windows, precluding a fair comparison with our method. In earlier experiments, these encoders performed uniformly poorly under the same objective: on AMIGOS, Criss-Cross Transformer reaches 34.60 F1 (Table 10 in the revised manuscript), BIOT 25.66 F1, and Medformer 21.85 F1, versus 66.23 F1 for AMIE. We emphasize that using a full long-duration ECG-EEG recording with its recording-level label as direct input constitutes a new protocol and provides a benchmark for full-length ECG-EEG modeling.

2. W2 + Q2 (Splits, subject IDs and recording selection)

Pretraining uses only training subjects. Validation and test subject IDs are listed in the table below. The three splits contain 469/80/80, 306/36/72, and 476/84/84 recordings (train/validation/test), respectively. For AMIGOS, we used 629 recordings and excluded 140, either because their durations are excessively long (846-1415s, roughly an order of magnitude longer than the 51-150s of other recordings) or due to missing data.

Since the three datasets contain only 40, 23, and 23 subjects and all suffer from class imbalance (severe in AMIGOS and DREAMER), random subject-independent splitting may leave certain classes absent from the validation or test subsets. We therefore follow [2] and [3] in assigning subjects to mutually exclusive training, validation, and test subsets. We have added these details to Appendix A of the revised manuscript.

DatasetValidation subject IDsTest subject IDs
AMIGOS6, 24, 26, 34, 392, 17, 21, 27, 38
DREAMER1, 43, 6, 10, 23
MPED1, 2, 34, 5, 6

3. W3 (Motivation against masked reconstruction)

We apologize for the incorrect citation in the Introduction of the original manuscript and for our inaccurate description of masked reconstruction; both have been corrected in the revised manuscript.

4. W4 + Q3 (Naming in Table 3 and the difference between Table 4 and Table 1)

The “All” rows at the bottom of Table 3 denote the concatenation of instance-level and global-level representations, the setting referred to as “Full” in the text; we have renamed these rows accordingly in the revised manuscript.

The discrepancy between Tables 4 and 1 stems from the protocol. The original Table 4 reports results with ECG and EEG concatenated before the encoder, without pretraining and with the concatenated input passed directly to the encoder, serving as a quick check of the GR design; Table 1 reports pretraining followed by fine-tuning. The table below reports the GR ablation with pretraining and fine-tuning and replaces the original Table 4 in the revised manuscript.

VariantAMIE F1(%)AMIE Acc(%)ModernTCN F1(%)ModernTCN Acc(%)PatchTST F1(%)PatchTST Acc(%)PatchTST+ F1(%)PatchTST+ Acc(%)OS-CNN F1(%)OS-CNN Acc(%)
w/o GR30.21±5.9846.39±4.2425.71±3.9844.45±6.7324.97±1.6839.45±7.8326.01±5.7339.45±2.5328.90±6.2546.11±4.21
w/o Refine33.36±5.1150.00±6.1331.41±2.5348.33±3.7626.67±6.4243.89±6.2624.62±4.5341.67±2.1929.44±7.0543.33±3.85
GR→TE34.05±3.2645.83±3.1031.82±4.9448.89±1.1628.67±4.3143.61±4.8726.79±3.8642.22±4.7732.21±6.9248.06±7.32
GR50.53±2.8963.61±2.0431.73±5.0550.00±7.0328.99±2.0244.44±3.2926.99±5.0840.83±8.2731.34±1.4548.89±2.55

5. W5 + Q4 (Figure 2 and the role of SGAM)

Figure 2 has been improved in the revised manuscript: the inputs ECG and EEG are now shown for SGAM, and the caption states that SGAM is used only during pretraining.

Reference:

[1] Ghallab et al., NeurIPS 2025.

[2] Xu et al., IEEE BIBM 2025.

[2] Zhong et al., IEEE BIBM 2025.

For W1 and Q1, I expected to see a comparison with state-of-the-art methods for cross-subject multimodal physiological emotion recognition, such as MDNet [1]. The Criss-Cross Transformer is still an architecture for EEG foundation models, not specifically developed for this task.

[1] Jia et al. Multi-level Disentangling Network for Cross-Subject Emotion Recognition Based on Multimodal Physiological Signals. IJCAI, 2024.

Thank you for your clarification. MDNet [1] takes short-segment physiological signals (6-second windows) as input. Our proposed HB-IGCL uses full long-duration ECG and EEG as input.

Thus, a direct performance comparison between MDNet and HB-IGCL is not feasible. The two tables below summarize the results of MDNet across three datasets, after being adapted to our experimental setup. Its overall performance is notably low. This outcome is predictable: MDNet is not purpose-built for full long-duration ECG and EEG inputs. The vast majority of other related models also split raw physiological signals into short segments, as stated in Figure 1 and the Introduction of the original manuscript.

F1 (%)

SeedAMIGOSDREAMERMPED
111132.2530.8225.39
222225.2828.1024.65
333329.3834.0228.25
444427.8231.2424.24
555525.0025.7126.19
Mean ± Std27.95 ± 3.0229.98 ± 3.1825.74 ± 1.59

Acc (%)

SeedAMIGOSDREAMERMPED
111132.0547.2235.71
222233.3351.3935.71
333334.6243.0655.95
444432.0538.8957.14
555534.6250.0032.14
Mean ± Std33.33 ± 1.2946.11 ± 5.1443.33 ± 12.16

We need to emphasize that the emotional information distributes sparsely and unevenly in long-duration physiological signals. The common pipeline of splitting signals into short segments and reassigning the recording-level label to every individual segment brings two critical flaws.

It may assign labels to segments that contain no emotional information (as illustrated in Figure 1 of the manuscript). It may also fragment continuous emotional segments.

This issue constitutes one of the core motivations of this work.

Criss-Cross Transformer in CBraMod [2] was originally developed for EEG decoding tasks. In prior work [3], the Criss-Cross Transformer has been adopted as the encoder for both ECG and EEG, to support bimodal emotion recognition. Table 10 reports the full performance results of the model proposed in [3].

References

[1] Jia et al. Multi-level Disentangling Network for Cross-Subject Emotion Recognition Based on Multimodal Physiological Signals. IJCAI, 2024.

[2] Jiquan Wang, Sha Zhao, Zhiling Luo, Yangxuan Zhou, Haiteng Jiang, Shijian Li, Tao Li, and Gang Pan. Cbramod: A criss-cross brain foundation model for eeg decoding. In International conference on learning representations, volume 2025, pp. 75310–75346, 2025.

[3] Youssef Ghallab, Omar Iraqy, Mohamed Kandil, Mohamed Ashraf, Saadeldine Eletter, Morougue Ghazal, Ayman Khalafallah, and Nagwa El-Makky. Leveraging foundational models and simple fusion for multi-modal physiological signal analysis. In NeurIPS 2025 Workshop on Foundation Models for the Brain and Body, 2025.

Thanks for the authors' response. I have decided to raise my score to 3.

Reviewer_85AhOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes HB-IGCL, a pretraining framework for long-duration ECG–EEG emotion recognition, which addresses the problem that segment-level supervision falsely assumes every segment to be emotion-bearing and that existing fusion strategies fail to exploit the synchronized local-to-global correspondence between the two modalities. It treats each recording as a bag of non-overlapping instances, replaces masked reconstruction with bidirectional cross-modal instance-global contrastive learning, and combines the multi-branch encoder, the GR module, and the SGAM lead-level spatial alignment. The paper further establishes systematic subject-independent benchmarks on three datasets, covering multiple encoders, multiple pretraining frameworks, five-seed statistics, and a full set of ablations, which fully demonstrate the effectiveness of the proposed objective.

Critical strengths

The paper addresses the real problem in long-term ECG-EEG emotion recognition, where the recording-level labels do not match the sparse emotional segments. It proposes HB-IGCL, replacing masked modeling with non-masked bidirectional instance-level comparison, and introducing SGAM to model the lead-level cross-modal structure. The experiments are comprehensive, covering three subject-independent datasets, five encoders, four pre-training frameworks, five sub-sets, and multiple ablation studies. The paper also reports the modality dominance related to the dataset, the superiority of Full-FT over Frozen-FT on MPED, baseline sensitivity, and research limitations. The innovative achievements are prominent, and the text is clear and easy to read. The experiments are sufficient and can clearly demonstrate the advantages of the innovation.

Critical weaknesses
  1. The core instance-global contrastive loss is closely aligned with existing multi-instance contrastive frameworks, such as MIL-NCE. The contribution needs to be repositioned.
  2. The headline claim is not supported by the statistical evidence and is not fully consistent with the reported results. Some Acc values are lower than the baseline, several F1 values have differences smaller than the standard deviation. The number of paired tests is insufficient, and the MPED of Frozen-FT on the AMIE is lower than the random initialization.
  3. The performance improvement mainly comes from the proposed AMIE. The independent contribution of the pre-training framework has not been isolated yet. The evidence for the SGAM spatial claim is weak.
Questions for authors
  1. The core instance-global contrastive loss is closely aligned with existing multi-instance contrastive frameworks, such as MIL-NCE. The contribution needs to be repositioned.
  2. The headline claim is not supported by the statistical evidence and is not fully consistent with the reported results. Some Acc values are lower than the baseline, several F1 values have differences smaller than the standard deviation. The number of paired tests is insufficient, and the MPED of Frozen-FT on the AMIE is lower than the random initialization.
  3. The performance improvement mainly comes from the proposed AMIE. The independent contribution of the pre-training framework has not been isolated yet. The evidence for the SGAM spatial claim is weak.
2 comments

Response to Reviewer 85Ah

We thank the reviewer for the careful and constructive assessment. We address each point below; new results will be added to the revised manuscript.

1. W1 + Q1 (Relation to MIL-NCE)

HB-IGCL and MIL-NCE both belong to the family of multi-instance contrastive learning. Their core “bag” definitions are fundamentally distinct. In MIL-NCE, a “bag” is constructed from candidate positive pairs. Each positive pair in MIL-NCE links one video clip to its multiple neighboring captions. In this work, a “bag” refers to a complete physiological signal recording. Each positive pair in our framework pairs a local ECG (EEG) instance with the global representation of full-length EEG (ECG). This distinction clearly separates our design from the MIL-NCE paradigm.

In the revised manuscript, we restate our contributions and introduce MIL-NCE in the Introduction.

W2 + Q2 (Statistical support and scope of the claims)

Our headline claim is not that every individual comparison is statistically significant, but that HB-IGCL achieves the highest mean F1 across all 15 fixed-encoder configurations. Table 1 reports mean ± standard deviation over five random seeds, a feasible and accepted protocol in this area [1]. We agree that paired tests are a valuable and widely used way to assess performance. At the same time, we would respectfully note that an aggregate claim over 15 configurations can also be supported by the consistency of mean rankings across seeds.

For MPED, Table 2 shows that Full-FT achieves the highest performance, with a substantial improvement over Frozen-FT, directly validating the efficacy of the proposed HB-IGCL pretraining framework. Although Rand. Init. exceeds Frozen-FT on MPED, we attribute this to two inherent properties of the dataset. First, MPED uses a 250 Hz sampling rate, twice the 128 Hz rate used for AMIGOS and DREAMER, endowing its raw recordings with markedly higher information density. Second, MPED’s raw recordings are far longer than most recordings in the other two datasets, enabling the extraction of a far larger pool of instances. Notably, Rand. Init. exhibits larger standard deviations across all metrics than both Full-FT and Frozen-FT. Corresponding clarifications are added to the revised manuscript.

W3 + Q3 (AMIE, pretraining and SGAM)

Pretraining frameworks can be compared column-wise in Table 1 (Section 4.3, “Performance comparison of pre-training frameworks”): for example, in the first column, HB-IGCL with the ModernTCN encoder achieves the best F1 on all three datasets, indicating that the HB-IGCL is the strongest. Encoders can be compared row-wise in Table 1 (“Performance comparison of encoders”): for example, in the last row, AMIE achieves the best performance, indicating that AMIE is the strongest encoder. Together, HB-IGCL with AMIE yields the best overall performance. Corresponding clarifications are added to the caption of Table 1 in the revised manuscript.

SGAM models inter-lead relations and thus inherently captures spatial information. Although we cannot directly prove that SGAM acquires spatial information, ablation results (Table 2) show clear gains: removing the SGAM lowers F1 from 66.23% to 60.32% on AMIGOS, from 50.53% to 48.89% on DREAMER, and from 84.66% to 83.85% on MPED. Identifying exactly what information a module captures is fundamentally an interpretability problem in deep learning and remains an open challenge.

Reference:

[1] Wang et al., ICLR 2025.

Thank you for the authors’ point-by-point response. The additional explanations helped me understand how the settings of HB-IGCL and MIL-NCE differ, but the methodological difference between their core objectives still needs further clarification.

Results from five random seeds and average rankings are not a substitute for paired tests of key comparisons. The reason why Frozen-FT performs worse than random initialization on MPED also remains insufficiently explained. In addition, the current comparisons do not clearly distinguish the respective contributions of AMIE and the pretraining framework, and the SGAM ablation alone is insufficient to directly demonstrate its spatial modeling mechanism.

I maintain my current ratings.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.