{"id":"e261b96b-68e3-474f-b987-dbea3b28f156","arxiv_id":"2505.20615","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A DCT-augmented EfficientNet trained on polysomnography-derived pseudo-images reaches 72.88% AUC for five-year incident hypertension in OSA patients, but the gain over baselines is not statistically supported.","lead":"A machine learning model converts sleep study signals into image-like grids and processes them with a JPEG-style frequency transform, predicting which sleep apnea patients may develop high blood pressure within five years. The headline 72.88 percent AUC is a modest gain over older models, but the paper lacks code, confidence intervals, and external validation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 3 Eq. (1) applies the DFT convolution theorem to DCT without justification; for standard DCT-II, element-wise multiplication does not implement convolution, so the DCT-based convolution framework, the paper's core novelty, is unsupported.","rationale":"Read in good faith: the paper aims to improve incident-hypertension prediction in OSA by inserting a DCT block into truncated EfficientNet. The central claim has two components: (a) the DCT layer improves prediction (best AUC 72.88 vs 71.64 without DCT); (b) this surpasses prior models. The reader's verdict is REJECT, with high correctness risk. I focused on the DCT convolution theorem because it is the only component that is internally checkable without data. The paper explicitly claims in the contributions and Section 3 that DCT replaces DFT for convolution, but the convolution-multiplication property is specific to the DFT (and other transforms under certain conditions). The DCT has a convolution theorem only for symmetric convolution, which is not what is implemented in a CNN. Therefore the theoretical foundation of the proposed layer is unsupported. This is a direct threat to the novelty claim, not just a missing baseline or CI. If the authors instead describe the layer as a spectral denoising/decorrelation transform, the model might still work, but the paper would need to be rewritten substantially and the current claims withdrawn. The concrete test is a simple numerical comparison that any reviewer can run, or an analytical derivation. If the concern lands, the paper cannot claim a 'DCT-based convolution framework.' This reinforces the reader's REJECT verdict, so no change.","tokens_in":7193,"tokens_out":8857,"duration_ms":89768,"concrete_test":"Implement the exact DCT block described (DCT-II, soft thresholding, IDCT) and the exact feature maps from a truncated EfficientNet. For a random filter w and feature map x, compute (a) y = conv(x,w) in the spatial domain, and (b) y' = IDCT(soft(DCT(x))·DCT(w)) in the DCT domain. Compare DCT(y) with DCT(x)·DCT(w) directly. If they are not equal (or proportional under the paper's normalization) for general x,w, the convolution theorem stated in Eq. (1) does not extend to the DCT, and the 'DCT-based convolution' claim is false. An analytical check: derive the DCT-II convolution-multiplication property; standard references show it requires symmetric convolution, not the operation implemented in the network.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 motivates the DCT layer by stating that convolution can be performed by element-wise multiplication in the frequency domain and that the DCT replaces the DFT to avoid complex values. This conflates the DFT convolution theorem with an analogous property that does not hold for the standard DCT. For DCT-II, Y[k]=W[k]X[k] corresponds to symmetric convolution only when both sequences are symmetrically extended; ordinary linear or circular convolution of two arbitrary sequences does not map to element-wise multiplication in the DCT domain. The paper neither states the required symmetry conditions nor shows that the 2D feature maps satisfy them. Consequently, the 'DCT-based convolution block' described as 'replacement of traditional convolutions' is not a convolution at all; at best it is a fixed orthogonal transform plus soft thresholding. This matters because the first listed contribution and Section 3's entire motivation for the layer are built on this false premise. The empirical AUC of 72.88% may still be achievable with a nonlinear DCT transform, but the paper's theoretical rationale for why the layer helps is invalid. Since the novelty claim rests on DCT-enabled frequency-domain convolution, this is a load-bearing flaw. Independent empirical concerns (no confidence intervals, post hoc depth selection, omitted SpO2 baseline) further weaken the superiority claim, but the DCT convolution error is the most direct threat to the paper's central methodological claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a deep learning pipeline for predicting five-year incident hypertension in patients with obstructive sleep apnea (OSA) from polysomnography (PSG) signals. The authors extract window-level features from multiple PSG signals, arrange them as 2D pseudo-images, and feed them into truncated 2D CNN backbones (EfficientNet, MobileNet, ResNet) pre-trained on ImageNet. Their main novelty is a DCT-based convolution block inserted at different depths of EfficientNet-B0, using soft thresholding instead of ReLU in the DCT domain. The best reported configuration (DCT@6) achieves 69.86% accuracy and 72.88% AUC, which the authors claim outperforms existing cSPPSG (71% AUC) and AHI (67% AUC) baselines. The paper also claims to be the first to jointly use all PSG signals for this prediction task.","tokens_in":7473,"tokens_out":3300,"duration_ms":36288,"significance":"If the proposed DCT-based method were valid and its performance confirmed, the paper would address a clinically important prediction problem and demonstrate a useful way to combine multi-channel PSG data in a transfer-learning framework. The authors also make a sensible design choice by extracting clinically meaningful features (respiratory events, HRV, statistical descriptors) before forming the pseudo-image, which could reduce data requirements. However, the central theoretical claim that DCT enables frequency-domain convolution is mathematically unsupported, and the empirical evaluation is weakened by the absence of confidence intervals, post-hoc configuration selection, and the omission of a cited SpO2-only model that reportedly achieves a substantially higher AUC. The paper does not provide code or a detailed experimental protocol, which limits reproducibility. The significance of the contribution, as currently framed, is therefore not established.","major_comments":[{"comment":"Eq. (1) states the DFT convolution theorem (Y = W * X) and then claims that the DCT can replace the DFT to avoid complex values. For standard DCT types (e.g., DCT-II), element-wise multiplication in the DCT domain does not implement ordinary linear or circular convolution of two arbitrary sequences; it corresponds to symmetric convolution only under specific symmetric extension conditions that the paper neither states nor verifies for the 2D feature maps. Consequently, the described 'DCT-based convolution block' is not a convolution replacement. At best, the layer is a fixed orthogonal transform followed by soft thresholding. Because the first listed contribution and the entire motivation for the DCT layer rest on this equation, this is a load-bearing mathematical error that invalidates the claimed methodological novelty.","section":"Section 3, Eq. (1)"},{"comment":"The comparison with existing methods omits You et al. [36], the SpO2-only model that the paper itself cites as achieving an AUC of 84.3% for the same task (incident hypertension in OSA patients). Including this baseline is necessary for a fair claim of 'surpassing existing models'. The proposed model's 72.88% AUC is considerably lower than this cited result. If the datasets or outcome definitions differ, the authors must state so explicitly and still discuss the comparison; otherwise, the superiority claim in the abstract and conclusion is misleading.","section":"Section 4.4, Table 4"},{"comment":"The final configuration (DCT@6) is selected after evaluating several DCT insertion depths and multiple window lengths (the text mentions windows from 9 to 60 minutes, with 10 and 60 minute results reported). The differences between the best configurations are small: DCT@5 AUC is 72.79% and DCT@6 AUC is 72.88%, compared with 71.64% for Feature EffNet-B0 at 10 minutes. No confidence intervals, bootstrap estimates, or significance tests are reported, and no multiple-comparison correction is applied to the selection across depths and windows. Without uncertainty quantification, the claim that deeper placement 'yields meaningful gains' is not supported. The authors should report confidence intervals for all reported AUCs and specify a pre-registered or corrected selection procedure.","section":"Section 4.3, Table 3"},{"comment":"The conclusion states that the method 'surpasses existing models such as cSPPSG and AHI', but the paper does not validate the model on an independent external cohort. The authors acknowledge the limited dataset size, yet this limitation is structural: without external validation, the generalizability claim remains unsupported, especially given the post-hoc configuration selection. An external validation (or at least a rigorous nested cross-validation with confidence intervals) is needed before the clinical-prediction claim can be taken seriously.","section":"Section 5 (Conclusion)"}],"minor_comments":[{"comment":"The text writes 'HR V metrics' in the feature list and later 'HRV metrics'; the former is a typo and should be corrected.","section":"Section 3, HRV metric description"},{"comment":"The text says DCT@5 gives AUC 72.29%, but Table 3 lists 72.79% for DCT@5. This inconsistency should be corrected.","section":"Section 4.3, text vs Table 3"},{"comment":"The phrase 'negative-frequency information' is inaccurate because the DCT does not produce negative frequencies; the DCT coefficients are real-valued and can be negative. The intended meaning is 'negative-valued coefficients', and the terminology should be revised.","section":"Section 3, soft thresholding discussion"},{"comment":"The description of the DCT block in Fig. 1 is incomplete: the figure is not explained in the text, and there are no details on the DCT type (e.g., DCT-II), the normalization, or how the 2D DCT is applied to the feature maps. This level of detail is necessary for reproducibility.","section":"Section 3, Fig. 1"},{"comment":"The experimental setup lacks many standard reporting details: number of subjects, class balance, hyperparameters (learning rate, batch size, epochs), and whether the 10-fold cross-validation was stratified. These should be added.","section":"Section 4"}],"recommendation":"reject","confidential_remarks":"The mathematical error in Section 3 is fundamental: the paper's core novelty is framed as DCT-based convolution, but the DCT convolution theorem does not apply to ordinary convolution without symmetry conditions. This cannot be fixed by local revision because it invalidates the central contribution as stated. The empirical evaluation would also need substantial additional work (confidence intervals, external validation, and a fair comparison including the cited SpO2 model) to support the claims. I recommend rejection, though the authors might consider resubmitting a revised manuscript that reframes the DCT layer as a feature transform and addresses the statistical concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. The genuinely new piece is narrow but real: the authors convert PSG-derived features into pseudo-images and test where a DCT plus soft-threshold block works when inserted at different depths of a truncated EfficientNet for five-year incident hypertension in OSA. The 72.88% AUC at depth 6 versus 71.64% for the best no-DCT model is the kind of result that would matter if it generalized. The paper is readable, the clinical question is worth asking, and the authors are upfront that the limited dataset may hurt generalizability.\n\nThe weaknesses, though, are not minor. Section 3 motivates the DCT block with Eq. (1), which is the DFT convolution theorem. That theorem does not transfer to the standard DCT-II. Element-wise multiplication in the DCT domain is equivalent to symmetric convolution only under specific symmetry conditions that the paper never states and that the feature maps do not obviously satisfy. So the 'DCT-based convolution block' is not a convolution replacement; it is a fixed orthogonal transform plus soft thresholding. That invalidates the first contribution and the stated reason the layer should help.\n\nThe empirical claims have similar problems. The best AUC comes from scanning window lengths and insertion depths on the same data, with no confidence intervals and no correction for multiple comparisons; a 1.2-point gain is well within plausible noise. The paper cites a SpO2-only baseline with 84.3% AUC [36] but leaves it out of Table 4, which makes the 'surpassing' claim look selective. Code and data are not provided, so the experiments cannot be reproduced or checked.\n\nI would not call this circular: no equation reduces to itself, and the selection problem is post hoc, not circular. But the paper as written is a preliminary engineering report, not a validated result. The DCT error is fixable in principle—rephrase the layer as a learned or fixed spectral transform, or justify symmetric convolution—and external validation plus confidence intervals would make the empirical claim testable. I would send it to peer review rather than desk-reject, because a competent referee can pin down whether the DCT block has real value; but I would not cite it in its current form.","headline":"A legitimate clinical prediction problem and an honest engineering effort, but the DCT-convolution rationale in Eq. (1) is wrong for the standard DCT and the 72.88% AUC claim lacks external validation.","tokens_in":7971,"tokens_out":4282,"would_cite":false,"duration_ms":42840,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","92C50"],"pacs":[],"model":"deepseek-v4-flash","headline":"DCT layer lifts hypertension-risk forecast to 72.88% AUC","keywords":["obstructive sleep apnea","hypertension prediction","discrete cosine transform","transfer learning","polysomnography","EfficientNet","frequency-domain convolution","soft thresholding"],"falsifier":"Take any two random sequences and compare the DCT of their convolution against the element-wise product of their individual DCTs: they differ unless special symmetry holds, which would undercut the theoretical rationale. Likewise, replacing the DCT2D block at depth 6 with an identity or a 1x1 convolution in the released code and re-running the same 10-fold cross-validation would show whether the reported 72.88% AUC depends on the DCT transform itself.","tokens_in":6976,"feed_emoji":"🩺","tokens_out":5824,"duration_ms":58386,"temperature":0.7,"pith_summary":"This paper tries to show that a deep network can predict whether a person with obstructive sleep apnea will develop hypertension within five years by looking at all their overnight sleep-study signals at once. The authors convert windowed features from EEG, ECG, respiratory, and oximetry channels into pseudo-images, then run them through a truncated EfficientNet whose convolutional blocks are augmented with a discrete cosine transform layer. They report a best area under the curve of 72.88% and accuracy of 69.86%, beating the cSPPSG composite phenotype at 71% AUC and the apnea-hypopnea index at 67% AUC. If true, this would make full polysomnography a usable input for long-term hypertension risk screening, not just for diagnosing apnea severity.","feed_headline":"DCT layer lifts hypertension-risk forecast to 72.88% AUC","feed_subtitle":"A neural net using every sleep-study channel at once beats AHI-based and composite phenotype scores.","key_machinery":"The carrying mechanism is the DCT-based convolution block: a real-valued orthogonal transform applied to 2D feature maps, followed by soft thresholding, $\\mathrm{sign}(x)\\max(|x|-\\tau,0)$, instead of ReLU. It is inserted after blocks 3, 4, 5, or 6 of a truncated EfficientNet-B0 that operates on a pseudo-image whose rows are extracted window features and whose columns are time windows. The DCT is meant to replace complex DFT-domain convolution via element-wise multiplication, while orthogonality decorrelates features and soft thresholding keeps both positive and negative high-amplitude coefficients.","core_discovery":"The paper's central claim is that inserting a DCT2D block deep inside a pretrained EfficientNet-B0 improves five-year incident hypertension prediction in moderate-to-severe obstructive sleep apnea. On the Sleep Heart Health Study cohort, the best configuration places the DCT block after the sixth building block and reaches 69.86% accuracy and 72.88% AUC, compared with 71% AUC for cSPPSG and 67% for AHI. The authors attribute the gain to transforming feature maps into a real, orthogonal frequency domain where soft thresholding prunes small noisy coefficients while retaining sign information that ReLU would discard. They also claim to be the first to feed all polysomnography signals jointly rather than a single channel or index.","pith_inferences":["Ablating the DCT2D block (replacing it with a 1x1 convolution or an identity at the same depth) would separate two explanations: exact frequency-domain convolution versus a learnable spectral feature transform; the paper does not run this control.","The same recipe could be applied to wearable multi-channel recordings for predicting other cardiometabolic outcomes, where sign-preserving frequency-domain nonlinearities may transfer.","Evaluating on an external OSA cohort would test whether the DCT layer learns transferable spectral patterns or patterns specific to the Sleep Heart Health Study dataset."],"forward_implications":["Clinicians could use the model's probability score, rather than the apnea-hypopnea index alone, to decide which OSA patients need closer blood-pressure follow-up over the next five years.","Combining all polysomnography channels beats single-index baselines, so future risk scores should stop reducing sleep studies to AHI or oxygen-desaturation indices.","Placing the frequency transform at deeper layers helps more than shallow placement, giving architects a concrete rule for where to insert spectral blocks in transfer-learned CNNs.","Soft thresholding in the frequency domain can be reused in other medical time-series tasks where keeping the sign of coefficients matters.","The feature-window pseudo-image format lets pretrained 2D image networks consume heterogeneous physiological signals without processing raw waveforms at full bandwidth."],"supporting_citations":[{"why":"Supplies the Sleep Heart Health Study design and the polysomnography data from which all features are extracted.","marker":"[11]"},{"why":"Provides the National Sleep Research Resource data commons that gives access to the SHHS signals.","marker":"[12]"},{"why":"Defines the cSPPSG composite sleep-pulmonary phenotype baseline that the proposed model is compared against and outperforms.","marker":"[23]"},{"why":"Prior feature-based pseudo-image machine-learning pipeline for incident hypertension prediction that this work extends.","marker":"[27]"},{"why":"Defines the EfficientNet architecture that is truncated and used with transfer learning as the CNN backbone.","marker":"[29]"},{"why":"SpO2-based random forest risk model with 84.3% AUC that motivates signal-only hypertension prediction in OSA.","marker":"[36]"}],"fun_headline_variants":["All sleep-study signals feed DCT model for hypertension risk at 72.88% AUC","DCT in deep net uses all PSG channels to hit 72.88% AUC for OSA hypertension","Frequency-domain DCT with all sleep signals predicts hypertension at 72.88% AUC","Deep net on all PSG signals with DCT transfer learning hits 72.88% AUC","All polysomnography signals + DCT improves 5-year hypertension prediction in OSA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that convolution can be carried out by element-wise multiplication in the DCT domain the same way it is in the Fourier domain; that identity only holds under special symmetric-convolution conditions the paper does not state, and the authors acknowledge that their limited dataset may limit generalizability.","fun_headline_variants_meta":{"raw":{"variants":["All sleep-study signals feed DCT model for hypertension risk at 72.88% AUC","DCT in deep net uses all PSG channels to hit 72.88% AUC for OSA hypertension","Frequency-domain DCT with all sleep signals predicts hypertension at 72.88% AUC","Deep net on all PSG signals with DCT transfer learning hits 72.88% AUC","All polysomnography signals + DCT improves 5-year hypertension prediction in OSA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00075,"raw_usage":{"total_tokens":3325,"prompt_tokens":918,"completion_tokens":2407,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":2290}},"tokens_in":534,"tokens_out":2407,"duration_ms":16706,"temperature":1.0,"reasoning_tokens":2290,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:50:33.618838+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any two random sequences and compare the DCT of their convolution against the element-wise product of their individual DCTs: they differ unless special symmetry holds, which would undercut the theoretical rationale. Likewise, replacing the DCT2D block at depth 6 with an identity or a 1x1 convolution in the released code and re-running the same 10-fold cross-validation would show whether the reported 72.88% AUC depends on the DCT transform itself.","supporting_citations":[{"cited_title":"& Others The sleep heart health study: design, rationale, and methods.Sleep.20, 1077-1085 (1997)","cited_arxiv_id":null,"evidence_quote":"Supplies the Sleep Heart Health Study design and the polysomnography data from which all features are extracted."},{"cited_title":"& Redline, S","cited_arxiv_id":null,"evidence_quote":"Provides the National Sleep Research Resource data commons that gives access to the SHHS signals."},{"cited_title":"& Sofer, T","cited_arxiv_id":null,"evidence_quote":"Defines the cSPPSG composite sleep-pulmonary phenotype baseline that the proposed model is compared against and outperforms."},{"cited_title":"& Prasad, B","cited_arxiv_id":null,"evidence_quote":"Prior feature-based pseudo-image machine-learning pipeline for incident hypertension prediction that this work extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SpO2-based random forest risk model with 84.3% AUC that motivates signal-only hypertension prediction in OSA."}],"review_version":1}