acceptodds
Withdrawn from ICLR 2027

Per-Recording Fidelity Guarantees for Wearable ECG Compression by Encoder-Side Verification

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In a wearable electrocardiograph (ECG), sending one bit over a short-range radio consumes about two thousand times the energy of one arithmetic operation, so battery life depends mainly on compression. Existing compressors are tuned once on a development set and applied to each recording the device later encounters, so fidelity can only be reported as an average, which conceals the recordings that were damaged. On the MIT-BIH ECG Compression Test Database the transform family achieves 23.15x while an automatic reader detects only 82.0% of the QRS complexes. We observe that the encoder retains the original signal, so it can decompress its own output before transmission, compare the beats detected in that reconstruction with the beats it started from, and reject any attempt that moves or loses a beat, retrying at a finer setting. Each recording is then compressed as much as it can tolerate, instead of as much as the least tolerant recording in the database allows, so the coder also achieves higher ratios. Over 162 recordings it obtains 14.81x with all beats preserved, compared with 6.49x at 99.3% beat retention for the published family, and it extends a 225 mA h cell from 304 to 631 days, with verification taking 9 percent of the power budget. Two further measurements limit what any coder can gain. Noise accounts for 87% to 96% of the bits used by a strong neural coder, and that coder is three orders of magnitude less energy-efficient than a linear predictor. Verifying fidelity inside the encoder allows ambulatory monitoring to be audited one recording at a time.

Primary area:
learning on time series and dynamical systems
Keywords:
Electrocardiography, data compression, wearable sensors, energy efficiency, signal quality.

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_C55dOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

For ECG compression, it is often difficult to decide the optimal compression ratio due to different compressibility of ECG signal. Hence, many works choose a conservative compression ratio in order to ensure signal fidelity for downstream tasks, which causes missing opportunities to achieve high compression ratio.

This paper’s main idea is that let the source to compress the signal as aggressive as possible, then decompress the signal and compare with it with the ground true signal at the source. If the decompressed signal meets beat detection criteria, the compressed signal can be sent, otherwise, the signal will be re-compressed with less aggressive compression ratio. The paper uses beat detection as a way of verification.

Critical strengths

None

Critical weaknesses
  1. The motivation, novelty, and achievements are overstated.
  2. Verification using “all beat preserved” is over simplified as a way for fidelity guarantee. ECG compression also requires clinically relevant waveform features are preserved.
  3. The evaluation protocol is not clearly presented. This causes concern if the comparison with baseline is fair.
  4. The memory requirements of the proposed encoder-side verification pipeline are insufficiently characterized.
  5. The battery-life estimation is not built on a sufficiently accurate energy model, hence, such analytical results are not sufficient for a top-conference like ICLR. I would suggest the authors to develop a prototype with a concrete hardware.
  6. It is not clear how much energy is used for performing compression, decompression, verification, and transmission.
Questions for authors
  1. Does compression preserve clinically relevant waveform features?

  2. Can the authors show how much energy is used for performing compression, decompression, verification, and transmission? As some decompressed signal cannot meet detection criteria, and it requires re-compression, how many signals on average require retries?

  3. What is the runtime memory footprint of the compressor, compression, decompression, verification? What is the peak RAM and processing latency? In particular, for the worst cases that requires retries.

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

This paper proposes a task-aware ECG compression framework for wearables. To prevent the loss of critical clinical events common in average-optimized methods, it adaptively verifies compression per-recording on-device. By ensuring compressed signals maintain original QRS detection accuracy, it practically balances high compression efficiency, low energy overhead, and clinical fidelity.

Critical strengths

This paper addresses the practical challenge of reducing transmission costs to improve battery life in wearable ECG monitoring. It proposes an intuitive encoder-side verification mechanism with a lossless fallback, achieving strong compression efficiency under strict detector-consistency constraints. The authors explicitly account for the verification's computational overhead, ensuring the framework is evaluated under realistic constraints. The authors provide code for their method and the reimplemented baselines, along with detailed evaluation settings, which supports reproducibility and future comparisons.

Critical weaknesses
  1. Limited comparison with recent work. Most baselines are classical ECG compression methods. While these are useful references, comparisons with more recent adaptive or learning-based methods are lacking, especially for an ICLR submission. This makes it difficult to assess the advantages of the proposed method over recent work.
  2. Limited generalizability. The noise estimation fails on a large portion of the PTB-XL recordings. It is unclear how well the method would work on these recordings, especially when lossless fallback is needed.
Questions for authors
  1. How does the proposed method compare with more recent adaptive or learning-based ECG compression methods under similar evaluation settings?
  2. How often is the monotonicity assumption violated in the bisection-based search, and how much does this affect compression performance?
  3. How does the method perform on recordings where noise estimation fails, particularly when lossless fallback is required?
Reviewer_AfV9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper argues that ECG compression on a wearable should be carefully considered because of the energy consumption associated with sending data compared to arithmetic on said data. The authors argue that quantizing data with a fixed approach must satisfy least-tolerable settings in a population. The authors' insight appears to be that the encoding mechanism holds the original data and can compress and decode its own input, and basically epmirically determine an acceptable setting based on whether or not a beat detector can match beat-for-beat with the original. The authors also report two measurements: noise accounts for 87–96% of the bits a strong convolutional coder spends, and that coder is ~3 orders of magnitude from energy break-even on a patch. A negative result is reported as well: a noise-proportional allowance did not protect a downstream diagnostic classifier better than a fixed allowance.

Critical strengths
  • The evaluation protocol is reasonable, and the technique is based on a reasonable insight. The authors' use of "beats retained" seems like a more desirable measurement than the use of PRD. This is desirable from the perspective of improving the overall rigor within the research community.
Critical weaknesses
  • Perhaps I missed something, but I could not see what type of beat detector was actually used. It is sort of vaguely described throughout the paper and evaluation. I also didn't really understand whether there was a ground truth sanity check. Just showing an analysis matching original quality with a proposed quality setting is not necessarily the same as demonstrating a match with respect to ground truth.
  • Somewhat related, but 160 ish second recordings are then used by the author to claim guarantees -- but I'm not sure this is quite convincing. Moreover, these aren't really guarantees because we don't know about the properties of the beat detector being used, etc.
Questions for authors
  • Is the noise budget applied in a way that is ecologically valid? What if the noise in actual radio communication does not match the type of noise estimated in the encoder's validation step?

  • What is the detector used here? Is D() a well-known function that I just missed when reading this manuscript?

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

Related papers

Discussion (0)

Sign in to comment.