{"id":"46cfcaa3-51bc-475c-bee0-2f858f86eddb","arxiv_id":"2411.15809","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Adding HODMD-derived mode images to the training set improved cardiac disease classification accuracy on held-out videos by up to 22 percentage points.","lead":"This paper combines a signal-processing tool called Higher Order Dynamic Mode Decomposition (HODMD) with a neural network to classify heart disease from ultrasound videos, and reports higher accuracy when the tool's extracted patterns are added to the training data. The authors present this as a new data augmentation method for medical imaging.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No control for training-set-size confound: the 42%→67% and 52%→74% gains on unseen data could be due to the 10% larger training set (7,000→7,700 images), not to HODMD modes.","rationale":"I agree with the reader's weakest-assumption identification. The central claim requires that the DMD modes—not the larger training set or the altered split—cause the accuracy gain. The paper provides no such control. The leaked 'testing' results should be set aside; on the only valid held-out data, the confound remains. A same-size original-image control is a simple, definitive experiment. Additional issues (manual mode selection, no error bars, no fixed selection criterion) further weaken the paper, but the dataset-size confound is the load-bearing one. Since the concern directly breaks the causal claim, the REJECT verdict stands unchanged.","tokens_in":10807,"tokens_out":4563,"duration_ms":43113,"concrete_test":"Train the same CNN (same architecture, optimizer, and hyperparameters from §2.2) on a control dataset of 7,700 original images—e.g., 110 frames per training sample instead of 100—split into 7,700/2,750/550 for training/validation/testing, and evaluate on the same 2,700-image testing I. If the control reaches approximately 67% (five-class) or 74% (four-class), the reported improvement is explained by training-set size rather than by HODMD. If the control is substantially worse, the confound is less threatening.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline result is that adding HODMD modes improves classification on truly unseen videos (Table 2: testing I five-class 42%→67%, four-class 52%→74%). But the comparison is between a model trained on 7,000 original images and one trained on 7,700 images (7,000 originals + 1,000 DMD modes). There is no control trained on 7,700 original images. With only 20 training videos per class, a 10% increase in training examples, or the different train/validation/test split (7,700/2,750/550 vs 7,000/2,500/500), could plausibly account for a large part of the gain. The improvement is therefore not attributable to the HODMD modes specifically. The paper's 'testing' results are leaked (images from the same videos as training), so the only valid comparison is testing I, and that comparison still carries this confound. Without a same-size original-image control, the central causal claim is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes using higher order dynamic mode decomposition (HODMD) to extract DMD modes from echocardiography videos and then using those modes as additional training images for a CNN that classifies five cardiac conditions: healthy, diabetic cardiomyopathy, obesity, myocardial infarction, and TAC hypertrophy. The authors compare two testcases: training the CNN on 7,000 original frames versus training it on 7,000 original frames plus 1,000 DMD-mode images, and they repeat the comparison after excluding the healthy class. Accuracy is reported on a test split of the original dataset and on 2,700 frames from 30 held-out videos (testing I). The paper concludes that combining original images with DMD modes improves classification accuracy by up to 22%.","tokens_in":11004,"tokens_out":7228,"duration_ms":63432,"significance":"The core suggestion, that modal-decomposition features can serve as a data augmentation source when medical training data are scarce, is interesting and the authors take the commendable step of including a held-out-video test set. The four-class experiment is a useful robustness check. However, the present evidence does not isolate the contribution of HODMD modes: the augmented model is trained on 10% more images than the baseline, and the internal test set is composed of frames from the same training videos. These are correctable design issues rather than flaws in the method itself, but until they are fixed the headline accuracy gains are not established. I also note that no code or data are provided, which limits reproducibility.","major_comments":[{"comment":"Testcase 01 is trained on 7,000 original images while testcase 02 is trained on 7,700 images (7,000 originals plus 1,000 DMD modes). The central comparison is therefore confounded: the higher accuracy on testing I (42% to 67% in the five-class experiment, 52% to 74% in the four-class experiment) could be produced by the additional 10% of training volume alone, without any DMD-specific effect. A control model trained on 7,700 original images, or a cross-validated comparison matched for training-set size, is required before the improvement can be attributed to HODMD.","section":"Section 3.1, Table 2"},{"comment":"The 'testing' set in dataset 01 is extracted from the same 20 videos per class that supply the training and validation frames, because 100 successive frames from each sample are later split into training, validation, and testing. Consecutive frames from the same echocardiography sequence are nearly identical, so the reported internal testing accuracies (76% to 94% and 79% to 98%) are inflated by frame-level leakage and should not be used as evidence of generalization. Only the held-out-video data (testing I) give a valid generalization estimate, and that estimate still suffers from the training-set-size confound described above.","section":"Section 3.1, Figs. 3c-3f and 4c-4f"},{"comment":"Testing I consists of 90 frames from each of only six videos per class, giving 30 videos in the five-class experiment and 24 in the four-class experiment. Reporting accuracy over 2,700 frames treats frames as independent observations when they are highly correlated within a video, and no confidence intervals or per-video accuracies are given. With this sample size, the 25-percentage-point improvement could be within run-to-run variability; repeated training runs and interval estimates are needed.","section":"Section 4, Table 2"},{"comment":"The selection of DMD modes for augmentation is not a fixed algorithm: the authors state that modes were 'plotted and well observed' and that 'most representative modes (clear patterns), with high amplitudes and low noise levels' were chosen, and the number of modes per sample (10) was set accordingly. This manual, class-aware selection is not reproducible from the manuscript, and the concluding claim in Section 5 that quality rather than quantity of modes is the correct criterion relies on an unreported experiment. The preprocessing should be specified as a reproducible rule, such as a tolerance-based automatic selection, or the subjective component should be justified as part of the method.","section":"Section 3.2"}],"minor_comments":[{"comment":"Table 1 is difficult to parse: the labels '4 classes classification' and '5 classes classification' appear as rows without values, and the healthy (H) row is listed even though Section 4.2 excludes healthy data from the four-class experiment. Please restructure the table to show the two experiments and their splits clearly.","section":"Table 1"},{"comment":"There are frequent typos and copyediting artifacts, including 'perfprmance' in the Figure 3 caption, 'di fferent' in the Introduction, 're presented' in Section 4.1, and 'the main scoop of this research' in Section 5, which should be 'main scope'.","section":"Throughout"},{"comment":"The terminology is inconsistent: Table 1 uses 'TAC hypertension' while the rest of the paper says 'TAC hypertrophy'; please standardize the disease name.","section":"Table 1 and text"},{"comment":"The hyperparameter tuning described in Section 2.2 reports only the learning rate; the search range and the final numbers of filters and dense units are not given, so the architecture choice is not fully reproducible.","section":"Section 2.2"},{"comment":"The HODMD tolerances epsilon_SVD and epsilon_DMD in Eqs. (5) and (8) are never given numerical values, and the index d for the DMD-d algorithm is not reported, so the number of retained modes and the temporal embedding are not reproducible from the manuscript.","section":"Section 2.1, Eqs. (5) and (8)"}],"recommendation":"major_revision","confidential_remarks":"This is an application paper with an interesting idea, but the current evidence is not sufficient because of the training-set-size confound and the frame-level leakage in the internal test set. A same-size original-image control and a subject-level split are essential before the central claim can be evaluated. If the authors can perform those experiments and report uncertainty measures, the paper could become publishable. I do not see a novelty disclosure issue; prior work by the same group is cited."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The central claim, that adding HODMD modes to the training set improves cardiac disease classification, is not supported by the current experimental design. The comparison is between a model trained on 7,000 original images and one trained on 7,700 images (7,000 originals + 1,000 DMD modes). There is no control trained on 7,700 original images, so the 42%→67% and 52%→74% gains on the held-out videos could be largely due to the extra 10% data rather than to anything specific about HODMD. The stress-test note is right: this confound is load-bearing.\n\nThat said, the paper is not without merit. The idea of using DMD modes as synthetic training exemplars is new, at least as far as the authors claim, and the external hold-out videos (testing I) are genuinely unseen and shared across both testcases, which is a good design choice. The writing is clear, the dataset is modest but real, and the appendix showing the DMD modes next to the original frames is helpful for seeing what the method extracts.\n\nThe soft spots, in order of severity:\n1. The confound above. It alone breaks the central claim as stated. A simple fix would be a control trained on 7,700 original images (or downsampling the augmented set to 7,000).\n2. The other \"testing\" results (76%→94% and 79%→98%) are on images drawn from the same videos as training, so they are leaky and not evidence of generalization. The paper does emphasize the unseen set, which is good, but the abstract highlights \"all the testcases\" without this caveat.\n3. The DMD mode selection is manual: \"plotted and well observed\" is not a reproducible criterion, and the appendix suggests the selection may be cherry-picked. No error bars or multiple seeds are reported for the CNN runs, so the 67% and 74% numbers come with unknown variance.\n4. The \"higher number of DMD modes\" experiment mentioned in the conclusions is not described, making the claim about quality-over-quantity hard to assess.\n\nFor readers, this paper is useful as a cautionary example of a confounded data-augmentation comparison and as a modest demonstration that DMD modes can carry disease-relevant information. The method itself deserves a serious look, but the manuscript needs a matched-size control and a reproducible mode-selection procedure before the headline claim can be taken seriously. I would not accept it as is, but I would encourage the authors to run the control and resubmit. A serious editor could send this to review with that request in mind.\n\nFor me: not citable in current form, but worth bringing to a reading group as a methodological case study.","headline":"The idea is real but the headline comparison is confounded: no control for the 10% larger training set means the claimed HODMD benefit on unseen videos is not established.","tokens_in":11599,"tokens_out":2042,"would_cite":false,"duration_ms":20500,"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":"A CNN trained on original echocardiography frames plus HODMD modes classifies unseen videos up to 22 points more accurately than one trained on frames alone.","keywords":["higher order dynamic mode decomposition","DMD modes","data augmentation","cardiac disease classification","echocardiography","convolutional neural network","deep learning","feature extraction"],"falsifier":"Train the identical CNN on 7700 original echocardiography images (the original 7000 plus 1000 additional original frames drawn from the same training samples) and test on the same 2700-image hold-out videos; if the accuracy matches or nearly matches the 67% and 74% obtained with DMD-mode augmentation, the modes are not the causal ingredient. A second check is to compare the confusion matrices on the hold-out set, since DMD modes specifically should alter which classes are confused.","tokens_in":10588,"feed_emoji":"🫀","tokens_out":8064,"duration_ms":66300,"temperature":0.7,"pith_summary":"This paper claims that higher order dynamic mode decomposition (HODMD), a data-driven modal analysis tool borrowed from fluid mechanics, can be repurposed as a data augmentation method for classifying cardiac disease from echocardiography videos. The authors extract dominant dynamic features—DMD modes—from each video, select ten modes per sample, and add these mode images to the training set of a convolutional neural network alongside the original frames. On held-out videos never seen during training, adding the modes raises five-class accuracy from 42% to 67% and four-class accuracy from 52% to 74%, an improvement of up to 22 percentage points. If that result holds, it would give medical-imaging groups with small, expensive datasets a way to improve classifier generalization without collecting more labeled videos.","feed_headline":"Train on DMD modes, gain 22 points in cardiac classification","feed_subtitle":"Adding modal-decomposition frames to echocardiography training lifts unseen-video accuracy from 42% to 67%.","key_machinery":"The engine of the method is the higher order dynamic mode decomposition (HODMD), an extension of dynamic mode decomposition that represents a video as a sum of modes $v(x,t_k)\\simeq\\sum_{m=1}^{M}a_m u_m e^{(\\delta_m+i\\omega_m)t_k}$, where $u_m$ are spatial modes, $\\omega_m$ frequencies, $\\delta_m$ growth rates, and $a_m$ amplitudes. Dimension reduction via singular value decomposition is followed by a DMD-d step based on Takens delay embedding, yielding the modes that are then treated as images. The paper uses these modes as additional CNN training samples; the mechanism proposed is that the mode images expose underlying motion and structure that raw frames hide, forcing the network to learn more transferable features.","core_discovery":"The central claim is that HODMD modes encode disease-specific dynamical structure in echocardiography sequences, and that training on these modes in addition to raw frames makes the CNN generalize to new subjects. The paper compares two training sets: dataset 01 with 7000 original images and dataset 02 with 7700 images comprising the same original frames plus 1000 selected DMD modes. Across all testcases the augmented model does better, most decisively on the unseen hold-out set, where accuracy moves from 42% to 67% for five classes and from 52% to 74% for four classes. The authors attribute the gain to the modes displaying enhanced patterns and details not perceptible in the original frames, and they report that increasing the number of modes beyond ten produced no additional improvement, which they interpret as evidence that mode quality, not quantity, is what matters.","pith_inferences":["Editorial inference: the reported comparison does not control for training-set size, so part of the gain may be a simple 10% increase in data; training a control model on 7700 original frames would isolate the contribution of the DMD modes.","Editorial inference: the mode-selection step is based on visual inspection of pattern quality, amplitude, and noise; a quantitative selection rule would make the augmentation reproducible and testable across datasets.","Editorial inference: since HODMD modes are computed from the same training videos, evaluating on videos from different centers or imaging systems would clarify whether the modes capture disease physiology or dataset-specific artifacts.","Editorial inference: the success of this approach suggests a broader class of spectral or modal feature augmentation for small medical datasets; comparing HODMD against PCA modes or wavelet features would reveal how much of the gain is specific to the dynamical decomposition."],"forward_implications":["On unseen hold-out videos, five-class accuracy rises from 42% to 67% and four-class accuracy from 52% to 74% when DMD modes are added.","Validation accuracy stays essentially flat (99% to 98%) while hold-out accuracy jumps, suggesting the modes improve generalization rather than training fit.","A small labeled dataset of 130 videos can be augmented with about 10% extra modal images without new data collection.","The same HODMD-extraction and augmentation pipeline could extend to other temporal medical image modalities, since the method is data-driven and does not require disease-specific handcrafted features.","The paper reports that mode quality rather than mode count drives the gain: more than ten modes per sample did not significantly help."],"supporting_citations":[{"why":"Defines the higher order dynamic mode decomposition algorithm used to extract DMD modes.","marker":"[17]"},{"why":"The authors' earlier work applying HODMD to heart-disease analysis; supplies the detailed feature-extraction step.","marker":"[18]"},{"why":"Introduces standard dynamic mode decomposition, of which HODMD is an extension.","marker":"[21]"},{"why":"Takens delay embedding theorem underpins the DMD-d step in HODMD.","marker":"[22]"},{"why":"Provides the Matlab implementation of HODMD and applications used to compute modes.","marker":"[23]"}],"fun_headline_variants":["HODMD data augmentation lifts cardiac CNN accuracy by 22%","Add dynamic modes to echo training, gain 22% heart disease accuracy","Dynamic mode decomposition improves cardiac classification by 22%","Augment echo frames with DMD modes: 22% better cardiac ID","Modal decomposition frames as extra training data lift heart AI 22%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the accuracy jump comes from the DMD modes themselves rather than from having 1000 extra training images: the paper compares 7000 original frames against 7700 frames (7000 original plus 1000 modes) without a control trained on 7700 original frames.","fun_headline_variants_meta":{"raw":{"variants":["HODMD data augmentation lifts cardiac CNN accuracy by 22%","Add dynamic modes to echo training, gain 22% heart disease accuracy","Dynamic mode decomposition improves cardiac classification by 22%","Augment echo frames with DMD modes: 22% better cardiac ID","Modal decomposition frames as extra training data lift heart AI 22%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000782,"raw_usage":{"total_tokens":3494,"prompt_tokens":1025,"completion_tokens":2469,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":641,"completion_tokens_details":{"reasoning_tokens":2378}},"tokens_in":641,"tokens_out":2469,"duration_ms":15657,"temperature":1.0,"reasoning_tokens":2378,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:51:49.567124+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the identical CNN on 7700 original echocardiography images (the original 7000 plus 1000 additional original frames drawn from the same training samples) and test on the same 2700-image hold-out videos; if the accuracy matches or nearly matches the 67% and 74% obtained with DMD-mode augmentation, the modes are not the causal ingredient. A second check is to compare the confusion matrices on the hold-out set, since DMD modes specifically should alter which classes are confused.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Takens delay embedding theorem underpins the DMD-d step in HODMD."},{"cited_title":"Le Clainche, J","cited_arxiv_id":null,"evidence_quote":"Defines the higher order dynamic mode decomposition algorithm used to extract DMD modes."},{"cited_title":"Groun, M","cited_arxiv_id":null,"evidence_quote":"The authors' earlier work applying HODMD to heart-disease analysis; supplies the detailed feature-extraction step."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces standard dynamic mode decomposition, of which HODMD is an extension."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Matlab implementation of HODMD and applications used to compute modes."}],"review_version":1}