{"id":"9fee1196-bd26-4366-b4b3-cf0696f65112","arxiv_id":"1908.03433","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A mixed DWT/DCT 2D codec for ECG signals achieves higher compression ratios than prior 2D methods on the MIT-BIH database for low distortion levels.","lead":"This paper presents an ECG compression codec that arranges heartbeats into a 2D array, then applies a wavelet transform along one axis and a cosine transform along the other. On the MIT-BIH database, it reports higher compression ratios than the same authors' 1D codec and than a 2D SPIHT benchmark at comparable distortion.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline improvement may be driven by a few highly regular records; without median or per-record significance tests, the claim of significant improvement is not established.","rationale":"The reader's verdict is CONDITIONAL, and I agree with that verdict. The reader identified QRS detection as the weakest assumption, but I see the more load-bearing issue as the statistical robustness of the headline claim itself. The paper reports only mean and standard deviation of CR over a distribution it admits is highly variable, with the benefit concentrated in records of regular morphology. This makes the mean an unreliable summary for the claim of general improvement. A re-analysis of the per-record data using medians and paired tests would directly settle whether the claim holds. The QRS detection concern is real but secondary: Pan-Tompkins is a mature algorithm and even if it occasionally misdetects, the effect on the mean CR is uncertain until tested. The availability of code makes the suggested re-analysis feasible, so the appropriate verdict remains CONDITIONAL pending that check.","tokens_in":4404,"tokens_out":5808,"duration_ms":65026,"concrete_test":"Using the released MATLAB code [6], compute per-record CR and PRDB for all 48 records for both the 1D and 2D codecs at the two operating points in Table II. Report the median CR, the proportion of records where 2D CR exceeds 1D CR, and the result of a paired Wilcoxon signed-rank test against (a) the 1D codec and (b) the fixed benchmark CRs of 30 and 20 from [5]. If the median 2D CR is not appreciably greater than 31 (or 30) and the majority of records do not improve, the mean-based headline claim is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim rests on Table II: at PRDB=6.82 the 2D codec reports mean CR 58 (std 63) against 31 for the 1D codec and 30 for SPIHT [5]. These means are not shown to be typical. Section V concedes that \"compressing in 2D is greatly beneficial for records of very regular morphology\" and that \"only some\" MIT-BIH records possess this trait, and Fig. 3 is a histogram of the per-record CR values. A mean of 58 with std 63 over n=48 records implies a strongly right-skewed distribution, so a few very regular records can dominate the average. No median, no interquartile range, no per-record comparison, and no paired significance test are reported. Without such evidence, the abstract's claim of \"significant improvements\" over benchmarks is not supported for the database as a whole; the improvement may hold only for a minority of records. This is the most load-bearing weakness because it attacks the headline claim directly, whereas the unvalidated QRS detector would only matter if detection failures are frequent enough to change the distribution.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a 2D ECG compression codec in which the 1D record is segmented into heartbeats, aligned into a 2D array, transformed with a 1D cdf97 DWT along the row direction and a 1D DCT along the column direction, then quantized, thresholded, and entropy-coded with Huffman coding. The method is tested on all 48 records of the MIT-BIH Arrhythmia database. The authors report that at a fixed distortion level (e.g., PRDB = 6.82) the 2D codec achieves a mean compression ratio of 58 with standard deviation 63, compared with 31 for their 1D codec and 30 for the SPIHT-based benchmark of Tai et al. They conclude that the 2D extension yields significant improvements over the 1D version and over prior 2D benchmarks, while acknowledging that the benefit is concentrated in records with very regular morphology.","tokens_in":4581,"tokens_out":3734,"duration_ms":44387,"significance":"If the claims are substantiated, the proposed codec is a simple and inexpensive way to exploit the regularity of ECG morphology for compression, with the notable strengths of being evaluated on the full MIT-BIH database and of providing MATLAB software for reproducibility. The mixed DWT/DCT transform is clearly described and the coding pipeline is standard. However, the headline claim of 'significant improvements' is currently supported only by mean values with very large dispersion, and the dependence on QRS detection is not validated. These issues are fixable, but they are central to the paper's main conclusion.","major_comments":[{"comment":"The abstract's claim of 'significant improvements' over the 1D codec and over 2D benchmarks is not statistically supported by the reported results. At PRDB = 6.82, Table II gives a mean CR of 58 with std 63 for the 2D method versus 31 with std 14 for the 1D method, over n = 48 records. Section V itself states that 2D compression is 'greatly beneficial for records of very regular morphology' and that 'only some' MIT-BIH records possess this trait, and Fig. 3 shows a strongly right-skewed distribution. A mean of 58 with std 63 can be driven by a small number of very compressible records, so the improvement may not hold for a typical record. The authors should report the median and interquartile range of per-record CR, present a paired per-record comparison between the 2D and 1D methods, and apply a paired significance test (e.g., Wilcoxon signed-rank on the per-record CR differences). The abstract and conclusions should be reworded to describe the improvement as holding for a subset of regular-morphology records unless the whole-database claim is justified by such evidence.","section":"Abstract; Table II; Section V"},{"comment":"The method's dependence on QRS detection is not validated. The paper states only that a MATLAB implementation of the Pan-Tompkins algorithm [8] was applied, but it reports no detection accuracy, no count of missed or false R peaks for the 48 records, and no sensitivity analysis. Because the 2D array is formed by aligning rows to detected R peaks, any detection failure changes the array content and hence the compressed file and reconstruction. The reported numerical results may therefore not be reproducible on other ECG databases or even on all MIT-BIH records if detection failures occurred. The authors should report per-record detection statistics and, if any records show misalignment, either exclude them with justification or analyze how the CR and PRD change. At minimum, the scope of the claims should be limited to signals for which reliable QRS detection is available.","section":"Section IV; Section II.A"}],"minor_comments":[{"comment":"The phrase 'World Hearth Organization' should read 'World Health Organization'.","section":"Section I"},{"comment":"The typo 'in oder' should be corrected to 'in order'.","section":"Section V"},{"comment":"The formatting 'MA TLAB' in Section IV should be 'MATLAB', and 'over-perform' in Section IV should be 'outperform'.","section":"Throughout"},{"comment":"The notation 'ℓi → ˜ℓi' is unclear; it would be clearer to state that the indices are reordered and the reordered sequence is denoted by tilde symbols.","section":"Section II.C"},{"comment":"The table reports no standard deviation for the benchmark method [5] because the original publication does not provide it; this should be stated explicitly in the text so that the large dispersion of the 2D results is not read as directly comparable.","section":"Table II"}],"recommendation":"major_revision","confidential_remarks":"The work is an incremental extension of the authors' own earlier 1D codec, and the novelty is modest but acceptable for a short communication. The central issue is statistical support for the headline claim; this should be addressable by adding per-record analysis and a paired test without new experiments. I would also gently encourage the editor to verify that the public MATLAB code reproduces the reported numbers, since reproducibility is one of the paper's strengths."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is a plausible incremental extension of the authors' own 1D ECG codec to 2D arrays, using a mixed DWT/DCT transform. It is clearly written, standard in its building blocks, and the full MIT-BIH results are reported with code available. But the headline claim of 'significant improvements' over the SPIHT benchmark is not actually established by the numbers as presented.\n\nWhat's new: the mixed transform itself — DWT along the heartbeat direction, DCT along the perpendicular — is a reasonable idea for leveraging beat regularity, and the paper shows visually that it concentrates energy better than a 2D DWT. That is a legitimate, if modest, contribution. The extension of the existing 1D pipeline to 2D is done carefully: quantization, zero-run indexing, Huffman coding, and the side information needed for reconstruction are all described in enough detail to replicate. Using the full 48-record database is good, and reproducing target PRD_B values to 0.01 across records is a nice touch.\n\nWhere the paper is soft: the central quantitative claim does not survive close reading. At PRD_B=6.82 the 2D codec's mean CR is 58 with a standard deviation of 63, versus 30 for the SPIHT benchmark. That standard deviation is larger than the mean and indicates a very right-skewed distribution. No median, interquartile range, per-record comparison, or paired significance test is reported. The paper itself concedes that only 'some' records benefit from 2D compression and shows a histogram of CR values that looks dominated by a handful of high-CR records. So the abstract's 'significant improvements' on average may be driven by a minority of regular beats, and the comparison to [5] is not statistically supported. This is the load-bearing weakness.\n\nA secondary concern is the unvalidated Pan-Tompkins QRS detection through a third-party MATLAB implementation. If beats are misaligned in some records, the 2D array is corrupted. The authors don't report detection failures or their effect on CR. Minor, since standard QRS detection on MIT-BIH is generally reliable, but it should be checked.\n\nWho this is for: researchers working on ECG compression and telemetry. It is a solid engineering paper that deserves a serious referee, but the revision should add per-record results, medians, and a proper paired comparison before the significance claim is accepted. I'd take a look at the revision if asked.\n\nRecommendation: send to peer review — the idea is real and the experiments are reproducible — but expect substantial revision.","headline":"Plausible 2D ECG codec extension with reproducible results, but the headline 'significant improvement' rests on a mean with enormous std and no significance test — deserves review, needs revision.","tokens_in":5127,"tokens_out":1678,"would_cite":false,"duration_ms":17103,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An ECG codec that applies a wavelet transform along aligned heartbeats and a cosine transform across them reaches mean compression ratio 58 versus 31 for the 1D version and 30 for a SPIHT benchmark at equal distortion.","keywords":["ECG compression","2D transform coding","mixed transform","discrete wavelet transform","discrete cosine transform","heartbeat alignment","lossy compression","SPIHT benchmark"],"falsifier":"Take one record with a high ectopic-beat count, run the full pipeline with the published software, and inspect the aligned array: if any R peak is not correctly aligned, the column-major zero runs shorten and the mean compression ratio of 58 at PRD_B=6.82 is not reproduced.","tokens_in":4178,"feed_emoji":"📈","tokens_out":9746,"duration_ms":89874,"temperature":0.7,"pith_summary":"An ECG record can be read as an image: the signal is cut into heartbeats, the R peaks are aligned, and the beats are stacked into a two-dimensional array. This paper claims that compressing that array with a wavelet transform along the beat direction and a cosine transform across beats outperforms both the same codec applied to the original one-dimensional signal and an established two-dimensional wavelet codec. At the same distortion, the mixed-transform codec reaches a mean compression ratio of 58 on the standard 48-record database, against 31 for the 1D version and 30 for the benchmark. The reason it works is that the mixed transform packs the important coefficients into a tighter region, so the entropy coder stores fewer nonzero values for the same quality. The paper also shows the method only helps above a certain distortion level; at very low distortion it tries to preserve sensor noise and stops being effective.","feed_headline":"Mixed wavelet-cosine codec doubles ECG compression at equal quality","feed_subtitle":"At equal distortion it reaches compression ratio 58 versus 31 for the 1D codec and 30 for the SPIHT benchmark.","key_machinery":"The central object is the mixed transform $B=\\hat W_{1r}\\hat C_{1c}A$, in which the 1D cdf97 discrete wavelet transform acts along the rows of the heartbeat-aligned array and the 1D discrete cosine transform acts along the columns. This object carries the argument because it determines where the large coefficients sit: in column-major order the significant entries cluster, so after quantization the vector of nonzero coefficients is shorter and its position deltas are smaller, making Huffman coding cheaper. The same choice fixes other design decisions, such as using six wavelet decomposition levels instead of the four that are optimal for the 1D codec.","core_discovery":"The paper's central claim is that the mixed transform—a cdf97 discrete wavelet transform on the rows of the heartbeat-aligned array and a discrete cosine transform on the columns, $B=\\hat W_{1r}\\hat C_{1c}A$—produces a coefficient layout better suited to entropy coding than either the 1D version of the codec or a 2D wavelet/SPIHT codec. With uniform mid-tread quantization, zero removal, sign separation, delta-index storage of nonzero positions, and Huffman coding, the 2D codec reports a mean compression ratio of 58 (standard deviation 63) at $PRD_B=6.82$ on all 48 records of the database, compared with 31 for the 1D codec and 30 for the benchmark; at $PRD_B=3.81$ the corresponding numbers are 26, 19, and 20. The authors attribute the gain to two effects: the mixed transform concentrates the significant coefficients, and six-level wavelet decomposition works better in 2D than the four-level optimum for 1D. They also report that the benefit is uneven, with large dispersion in compression ratio because only records with regular beat morphology compress well, and that the method is not effective below $PRD\\approx 0.4$, where reproducing the signal means reproducing sensor noise.","pith_inferences":["A practical encoder could classify beats by morphology and route regular beats to the 2D path and irregular beats to the 1D path; this would likely shrink the large dispersion (standard deviation 63 at mean compression ratio 58) that the paper reports.","The fixed transform orientation—wavelet on rows, cosine on columns—is one choice; adapting the orientation per record, or choosing which dimension gets the wavelet by measuring coefficient concentration, is a natural testable improvement not explored in the paper.","The same segmentation-and-alignment construction applies to other quasi-periodic biological signals, so the codec could transfer to pulse oximetry or respiratory waveforms with only the beat detector swapped out."],"forward_implications":["At the same distortion level, the 2D codec roughly doubles the mean compression ratio of the 1D version on regular-morphology records, so long-term ECG storage and telemetry can carry more data per bit.","The single quantization parameter $\\Delta$ lets an encoder target a desired PRD without changing the decoder, which simplifies rate control in practical devices.","Because the codec operates on raw data, it fails at very low distortion (PRD below about 0.4), where it would need to reproduce sensor noise; usable operation is limited to moderate and higher distortion levels.","The method requires QRS detection and heartbeat alignment, so its compression gain is conditional on beat regularity; records with irregular morphology show much lower compression ratios and larger dispersion."],"supporting_citations":[{"why":"The 1D compression strategy that the 2D extension builds on and directly compares against in Tables I and II.","marker":"[1]"},{"why":"The 2D wavelet/SPIHT benchmark that supplies the baseline distortion levels and compression ratios reproduced in Table II.","marker":"[5]"},{"why":"The standard 48-record ECG database that provides the test signals for every reported numerical result.","marker":"[7]"},{"why":"The QRS detector implementation used to segment and align heartbeats into the 2D array.","marker":"[8]"},{"why":"The QRS detection algorithm that the implementation follows and that supplies the R-peak positions needed for the 2D construction.","marker":"[9]"}],"fun_headline_variants":["Mixed wavelet-cosine transform lifts ECG compression ratio to 58","2D wavelet-DCT beats SPIHT and 1D ECG compression","ECG codec: wavelet rows, cosine columns, ratio 58 vs 30","Heart signal 2D codec doubles compression over SPIHT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the heartbeat detector finds every R peak correctly; a single misaligned beat corrupts a row of the 2D array and the reported compression-ratio gains would not be reproduced.","fun_headline_variants_meta":{"raw":{"variants":["Mixed wavelet-cosine transform lifts ECG compression ratio to 58","2D wavelet-DCT beats SPIHT and 1D ECG compression","ECG codec: wavelet rows, cosine columns, ratio 58 vs 30","Heart signal 2D codec doubles compression over SPIHT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001007,"raw_usage":{"total_tokens":4234,"prompt_tokens":901,"completion_tokens":3333,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":3255}},"tokens_in":517,"tokens_out":3333,"duration_ms":23228,"temperature":1.0,"reasoning_tokens":3255,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:12:41.407033+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take one record with a high ectopic-beat count, run the full pipeline with the published software, and inspect the aligned array: if any R peak is not correctly aligned, the column-major zero runs shorten and the mean compression ratio of 58 at PRD_B=6.82 is not reproduced.","supporting_citations":[{"cited_title":"Rebollo-Neira, ``Effective high compression of ECG signals at low level distortion'', Scientific Reports, 9 , No 4564 (2019)","cited_arxiv_id":null,"evidence_quote":"The 1D compression strategy that the 2D extension builds on and directly compares against in Tables I and II."},{"cited_title":"on Biomedical Engineering , 52 , pp 999 -- 1008 (2005)","cited_arxiv_id":null,"evidence_quote":"The 2D wavelet/SPIHT benchmark that supplies the baseline distortion levels and compression ratios reproduced in Table II."},{"cited_title":"Cohen, I Daubechies, and J","cited_arxiv_id":null,"evidence_quote":"The standard 48-record ECG database that provides the test signals for every reported numerical result."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The QRS detector implementation used to segment and align heartbeats into the 2D array."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The QRS detection algorithm that the implementation follows and that supplies the R-peak positions needed for the 2D construction."}],"review_version":1}