- Rating
- 2(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
DANCE introduces a ssl loss that makes the similarity between electrode embeddings match their functional connectivity defined by amplitude-envelope correlation, and tests it on Brain Treebank speech decoding and a private four-patient iEEG dataset for sleep staging and seizure detection.
Critical strengths
1.This paper studies an important problem
2.The motivation is natural and straightforward
Critical weaknesses
-
The experiment results are not impressive. But I do appreciate the author's honesty in providing the p value, and i like how author expose that , random initialized model or only hand power is doing very well. But the p-value and the good performance of the very simple model and the random-initialized, no pretraining model just show that the improvements are not that significant. Thus, it does not support the effectiveness of the proposed ssl loss
-
This paper advertises itself as 'foundation model for iEEG', but really its main contribution is the unsupervised loss objective of learning the amplitude-envelope correlation defined channel relationship. And in the evaluation pipeline, each dataset is pretrained on itself, then linear probe fine tuen, so it is really not in a foundation model fashion
Questions for authors
please see the weakness
2 comments
Thank the author for the last-minute response, I really appreciate your effort, and raised the score to 2
Response to Reviewer TgCN
We thank the reviewer for the assessment. New experiments are in three tables of the general response: Table R1 (Brain Treebank, additional pretrained encoders), Table R2 (transfer to the chronic cohort) and Table R3 (transfer to another public dataset, Omni-iEEG).
1. Significance of the improvements. The gain of DANCE over the same encoder from random initialisation is significant on speech in the submission (+0.024, p = 0.031; with the evaluated patient excluded from pretraining, +0.030, p = 0.031), and it now replicates with two newly pretrained encoders (+0.039 and +0.037, both in 7/7 subjects, p = 0.016; Table R1). With seven subjects, improvement in all seven gives p = 0.016, the smallest attainable value. On the chronic cohort, pretraining also improves over random initialisation in 4/4 patients on both sleep staging and IED rate (Table R2). On sentence onset the gains are not significant because the task leaves little headroom: all models, including a linear model on band power, lie within about 0.02 of each other (0.846 to 0.867).
For the same reason the absolute gains on Brain Treebank are modest: BaRISTA itself is within 0.02 of a linear model on band power on both tasks. On this benchmark DANCE is comparable to BaRISTA on average without using electrode coordinates, which BaRISTA requires as input; on the chronic cohort it outperforms BaRISTA (Table R2), and on Omni-iEEG it is above BaRISTA's temporal encoder (Table R3).
2. The "foundation model" framing and pretraining on each dataset. On the framing, we agree: the contribution is the pretraining objective, and we have replaced "foundation model" with "self-supervised pretraining objective" throughout. On the evaluation, the submission already included an experiment in which the evaluated patient is excluded from pretraining (last row of Table 1 and the leave-one-patient-out appendix), with a gain of +0.030 (6/7 subjects, p = 0.031). Since this is transfer across patients, not across datasets, we added two experiments in which an encoder pretrained on Brain Treebank only is applied to other datasets:
We hope these additions address the concern that the results do not support the objective, and we would be glad to answer any further question.