acceptodds
Withdrawn from ICLR 2027

RespFocus: Learning Temporal Evidence for Weakly Supervised Respiratory Sound Classification

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Respiratory auscultation provides a low-cost means of assessing pulmonary function, yet transient crackles and sustained wheezes can be difficult to recognize in noisy clinical recordings. Conventional deep learning classifiers often operate as black boxes, offering limited insight into the temporal acoustic evidence supporting their predictions. We present RespFocus, a weakly supervised framework that uses multiple instance learning (MIL) to learn inspectable temporal evidence from cycle-level labels and integrate it into classification. We also introduce PedResp2026, a multicenter pediatric respiratory sound cohort collected in routine clinical environments with adjudicated cycle-level annotations. In classification experiments, RespFocus achieves four-class Scores of % on public ICBHI 2017 and % on PedResp2026. Evidence diagnostics and matched temporal interventions support the functional contribution of the learned evidence to model predictions. Together, these findings highlight RespFocus's potential as a tool for automated respiratory sound diagnosis.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Respiratory sound classification, Weakly supervised learning, Multiple instance learning, Audio representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_XspfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The RespFocus paper deals with a weakly supervised respiratory-sound classification framework that uses multiple instance learning (MIL) to learn temporal evidence from respiratory cycle-level labels rather than requiring event-level annotations.

It combines OPERA-CT and AST audio representations, uses shared abnormal-presence evidence to guide separate crackle and wheeze branches, and preserves their differing temporal patterns through crackle max pooling and wheeze top-k pooling. The subtype decisions are then composed into four categories: Normal, Crackles, Wheezes, and Both

The paper also introduces PedResp2026, a multicenter pediatric cohort with 493 recordings from 137 children across seven hospitals, using expert respiratory-cycle labels and patient-disjoint evaluation (i.e. patients used during training, are not used during evaluation)

Critical strengths
  • RespFocus analyzes each annotated respiratory cycle as a MIL bag and learns which temporal tokens within that cycle provide evidence for abnormal sounds - without requiring event-level annotations. Each prediction is accompanied by subtype-specific crackle/wheeze evidence aligned to the original cycle interval. This helps to analyze the decision of the model - as shown in the Figs.

  • RespFocus achieves the highest reported mean four-class classification Score among the compared methods on both benchmarks: 69.55% on ICBHI 2017 and 72.91% on the new PedResp2026 cohort. These results improve over the strongest baselines

Critical weaknesses
  • RespFocus combines two pretrained feature streams, cross-attention of both, separate convolutional subtype branches, different pooling rules, ..., and independently selected thresholds. The ablation demonstrates incremental gains, but it does not fully isolate the contribution of all training objectives and modeling choices. The motivation for the particular design choice/composition of the model architecture is not always clear/supported by ablations (there is Table 2 - but it still lacks a clear motivation why particular stages are useful. Also there are many hyperparameters - not all of them seem to be set via the validation fold. Some ablations might be useful.

  • The final four-class label is obtained by independently thresholding crackle and wheeze probabilities. Threshold pairs are selected through validation grid search, and the Both label is assigned whenever both decisions are positive. This approach is sensitive to calibration and threshold choice and is not optimal

  • The paper argues that RespFocus learns interpretable temporal evidence, but it never evaluates whether the highlighted regions actually coincide with expert-annotated crackle or wheeze events.

Questions for authors

Is the PedResp2026 publicly avaliable?

Does the method work without annotated cycle intervals? Do the comparison methods in Table 1 require the cycle labels?

Adventitious lung sound classification is interesting, but would it be not much more valuable to have disease classification?

Why are the two subtype probabilities combined via thresholding (determined by validation data) to form the final output. Why is this not simply learned? How do the results change if you learn this combination?

What is exactly meant with evidence pathway in Table 2 - is it the refinement?

The optional text interface is entering towards the end of the paper. This is a bit confusing - maybe consider to remove it from the paper.

Remark: There is other work on respiratory disease classification using MIL available (a short version appeared at IEEE EMBC). The aim there is disease classification. It might be worth to mention it. Nguyen, T. T. K., Nguyen, M. D. N., Nguyen, S. V., Tran, T. T. N., Ma, N. T. H., Quach, H. T., Pham, A. L., & Pernkopf, F. (2026). “Lung-Sound Respiratory Disease Classification via Multiple-Instance Learning.” IEEE Access, 14, 41333–41349. DOI: 10.1109/ACCESS.2026.3669914.

Reviewer_3FsVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RespFocus use multiple-instance learning (MIL) to classifies respiratory cycles under weak supervision using only cycle-level labels. The method fuses OPERA-CT and AST features by cross-attention, then a shared abnormal-presence head reweights the time tokens. The method achieves higher performance compared to baselines on both a public and private dataset.

Critical strengths
  1. good research question: making respiratory sound classifiers expose the temporal evidence behind each prediction, which matters for clinical interpretability.
  2. the new pediatric dataset will be a valuable resource to the community
Critical weaknesses
  1. No significance test is reported, and RespFocus has much larger seed variance than the baselines.The margin over the second best method is smaller than RespFocus's own seed SD in ICBHI, for both 4 class and 2 class, also on PedResp2026 for 2 class.
  2. The core claim, temporal evidence, is never checked against event-level ground truth, although public datasets with event annotations exist (SPRSound, HF_Lung). The qualitative example in figure 7 also do not show how the evidence is helpful or useful. The model seem to have similar behavior for correct or incorrect predictions. It is unclear whether the model works as designed.
  3. The LLM report module is not quantitatively evaluated and adds no technical contribution.
Questions for authors
  1. In Figure 4, zeroing the entire cycle changes the subtype probabilities by less than 0.01? Does this mean predictions are driven mainly by context outside the cycle? Please report deletion and retention averaged over the test set.
  2. Do the baselines get the same validation-tuned, Score-maximising thresholds? What does RespFocus score with 0.5 thresholds?
  3. Please give paired 95% confidence intervals against the baselines.
  4. PedResp2026: What is the inter-annotator agreement? How were cycles chosen for annotation? The data has only about 2.3 annotated cycles per recording, each at least 30s long.
Reviewer_HVULOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Summary & Contributions

The paper focus on respiratory sound analysis from auscultations. Specifically, it focuses on the identification of crackles and wheezes (or both). It presents a particular architecture, a multiple instance learning (MIL) training paradigm, and a new dataset on top of a standard, open-source dataset (ICBHI).

Reviewer Recommendation

I have serious issues with the presentation of this work. I will outline concrete problems in the respective weakness section, but even as an expert (see self-evaluation), this paper is hard to follow and filled with jargon that I've never encountered before, especially in the presentation of the method. This makes it very hard to judge what the algorithm is supposed to be doing. Even though the reported results outperform the chosen baselines, I'm not sure what to make of them.

For this reason, I am scoring this paper as a clear rejection, though I am open to change my mind after rebuttal.

Reviewer self-evaluation

I have been working on audio for the last 8 years and have spent considerable time on respiratory diseases and audio foundation models. My specialty is not on respiratory sounds, but rather on speech. Nevertheless, I am well familiar with this research area and confident in my review.

Critical strengths

New dataset

One major advantage is the new dataset, which will be published along with the paper (with controlled access, given the privacy of the data). This is a big advantage.

However, it is very hard for me to justify why this should be an advantage for a main track paper of ICLR. It could go under a "Datasets" track and the data could be published as a standalone contribution, if it is to be made a major contribution.

Good benchmark results

Though not all comparisons are clear or fair, the results on ICBHI appear to be competitive, which is always interesting for a part of the community. I personally am not so interested in that, especially if I have serious trouble understanding what the method is doing and why (see below), but I think it's fair to say that other members of the community don't appear to mind this

Critical weaknesses

Presentation

As I said previously, there are numerous things that make it very hard for me to follow this paper, to the point where I was almost to recommend it's flat-out rejection due to presentation issues. I will try to give a non-exhaustive list of the main problems

Method

Why dual-stream

The authors go straight into a dual-stream architecture, fusing embeddings from AST and OPERA-CT. It never becomes clear why this should be done or what happens if each stream is examined in isolation. I guess this last part might happen in table 9, but it's impossible since there is only a cryptic description and the caption is not informative. (Also, this is packed in the Appendix, with neither the reviewer nor any future readers should be obliged to read).

However, in that table the pooling is also not ablated, so it's not clear if "Acoustic view dual stream with tSCA (Baseline)" is with pooling or without and what happens if you add the pooling to the one thing or the other --- in short, quite sloppy and not clear for the reader at all.

In particular, what is the best "full" pooling or whatever completely ablated and optimized pooling on OPERA-only and AST-only. In short -- how important is it to fuse the two streams? Because if it only works when fusing but not when not fusing, then this is weird and it gets us into that whole sort of "let's stack more encoders until it works better than the previous SOTA" research, which is not particularly interesting

Temporal-spectral cross-attention?

Why this title when the equations imply that good old "Attention is all you need" MHA is applied? If this is the simple MHA, then call it MHA (and do not imply that this is a new invention)

Gating (3.2)

This section is completely opaque to the reader (esp. because core parts are offloaded to the appendix), but I'll try my best:

  1. There is some "auxiliary abnormal cue" which I have no idea what it is
  2. In any case, this cue is turned into a logit (details in the appendix; be aware that this is something a reviewer should not be obligated to check)
  3. There appears to be a linear layer applied on each token, then another linear layer mapping it to 4 dims (the number of classes), then an exponent, then a log, then a weighted addition adding the logged-exponent-linear to the other linear (with a magic number determining the weight factor), then, finally, a sigmoid
  4. (Back to the main paper!)
  5. This is now used as elementwise product to gate some of the tokens

I guess this is a somehow reasonable "neural block" to apply but I take issue with the following two things:

  1. Its presentation is unnecessarily split between appendix & main text
  2. There are too many bizarre, unjustified, slightly anthropomorphizing terms. E.g. the authors arbitrarily name stuff as "abnormal channel index" or "auxiliary abnormal cue", without every explaining what is abnormal, why it is abnormal, how this was measured -- or who was auxiliary to whom

Sorry, but this level of presentation is not adequate in my eyes

Results

Without going into many details, many of the results have this back-and-forth between the appendix and the main text. I understand that the authors have a lot to show, but if the 9 pages of the main track do not suffice, then perhaps they have chosen the wrong venue.

Time-series baselines

The time-series baselines are never properly introduced (what are they? what features are they based on? are they made for audio signals? highly unlikely...). Their results are presented in an appendix and are substantially below the tested method. what is the reader supposed to make of this?

(I guess this can simply be removed)

Conclusion on presentation issues

I will not go deeper into details, but this paper is extremely hard to follow. This makes it very hard for anyone to either reproduce this work or to figure out why things worked the way they did. The authors are strongly advised to revise, both in case of acceptance and of rejection (I guess they'll ultimately resubmit and this paper will eventually see the light of day..)

Too much unconnected stuff

The authors present:

  1. A new dataset (the most valuable contribution).
  2. A new method with a. A dual-stream encoder b. Some gating function c. Multiple instance learning with different tweaked pooling strategies per abnormality type
  3. Baselines from a different disciplines
  4. An appendix packed with a method to map the output of the model to text via an LLM (which I so far ignored)

I guess the versatility of the authors must be applauded.

On the other hand, I am personally missing a "story" out of this paper. What does it mean? Why should the reader care? Perhaps this is a personal preference, but I have would have preferred less diversity and more focus.

Questions for authors
  • Some clarification regarding the gating function might help (see above for details)
  • See above for details on experimental comparisons

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

Related papers

Discussion (0)

Sign in to comment.