{"id":"6b3a30b8-8731-418a-96b4-8ddf1310fcf3","arxiv_id":"2607.07683","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":7,"one_line_summary":"An end-to-end YOLOv11-based pipeline digitizes paper ECG images into calibrated 12-lead signals on CPU-only hardware in under 30 seconds and classifies myocardial infarction with up to 95.5% accuracy on PTB-XL and 88.9% on ECG-Matrix.","lead":"This paper builds a CPU-only pipeline that converts photos of paper ECG printouts into digital 12-lead signals and screens for myocardial infarction. A smart generalist might read it because it targets a real bottleneck: millions of archived paper ECGs in low-resource clinics are currently inaccessible to automated cardiac diagnosis.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"Headline 95.51% accuracy comes from a synthetic-to-synthetic digitization pipeline (train and test on ECG-Image-Kit renderings of PTB-XL), not from real paper ECGs; the only real-world test (ECG-Matrix, 88.89%) lacks signal-level ground truth and trains classifiers on the same digitized distribution","rationale":"The reader's verdict of CONDITIONAL with MODERATE confidence is appropriate. The identified concern about the synthetic-to-real domain gap is the correct load-bearing issue, and I partially agree with the reader's framing. My sharpening is that the concern has two distinct facets the reader did not fully separate: (1) the headline 95.51% accuracy is synthetic-to-synthetic and does not test real-world generalization at all, and (2) the ECG-Matrix 88.89% result, while genuinely testing real-world digitization, is structurally insensitive to systematic artifacts because classifiers are trained and tested on the same digitized distribution.\n\nThe paper has real strengths: code and pretrained models are publicly available, the engineering is sound, patient-wise splits are used throughout, and the SHAP attributions are physiologically plausible. The MLP architecture inconsistency (Methods say one hidden layer of 100 neurons; Supplementary says two layers of 100 and 50; main text mentions 50 PCA components not described in the architecture) is a documentation issue but does not affect the central claim.\n\nThe CONDITIONAL verdict is correct because the paper makes a legitimate engineering contribution but overstates the evidence for real-world generalization. The abstract implies the combined accuracy numbers validate the full pipeline, when in fact only the ECG-Matrix result tests real-world digitization, and it does so only indirectly. A prospective validation on real paper ECGs with signal-level ground truth (as described in the concrete test) would be needed to move from CONDITIONAL to ACCEPT. The paper itself acknowledges this need in the Discussion, which is appropriate, but the abstract and conclusions could be clearer about which evidence supports which claim.","tokens_in":36074,"tokens_out":3672,"duration_ms":237963,"concrete_test":"Print 50–100 PTB-XL waveforms onto physical paper using a standard ECG printer, photograph them with a smartphone under realistic clinical conditions (variable lighting, angles, partial occlusion), then run them through the digitizer and compare reconstructed signals to the known PTB-XL ground truth using the same metrics (SNR, Pearson r, RMSE). If the signal-level metrics drop substantially below the synthetic-to-synthetic values (e.g., r < 0.6 or failure rate > 25%), the synthetic-to-real domain gap is larger than the indirect ECG-Matrix classification results suggest, and the generalization claim would need qualification. If metrics remain comparable, the generalization claim is strengthened with direct signal-level evidence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim has two parts: (1) a digitizer trained on synthetic PTB-XL images generalizes to real hospital paper ECGs, and (2) the full pipeline achieves 95.51% MI accuracy on PTB-XL and 88.89% OMI accuracy on ECG-Matrix. These two results carry very different evidential weight.\n\nFor PTB-XL (95.51%): The digitizer is trained on ECG-Image-Kit renderings of PTB-XL waveforms and evaluated on held-out renderings from the same generator (Table S1C, 1,600 images). Classification is then performed on these synthetic-to-synthetic digitized signals. This pipeline never touches a real paper ECG, so the 95.51% headline accuracy does not test real-world digitization generalization at all—it measures classification on signals that passed through an in-distribution synthetic digitizer with moderate fidelity (SNR=4.54 dB, r=0.806, 11.94% failure rate).\n\nFor ECG-Matrix (88.89%): This is the only test of real-world generalization. However, the classifiers (e.g., Rocket) are trained directly on the digitized ECG-Matrix signals with patient-wise splits (Table IIb: 1,997 train / 666 test beats from 319/107 patients). Because train and test signals come from the same digitization distribution, any systematic artifacts introduced by the synthetic-to-real domain gap would be present in both splits. The classifier could learn to exploit or tolerate these artifacts without them affecting accuracy. Thus, good downstream classification does not establish that the digitizer faithfully reconstructs real paper ECG morphology—it only shows the digitized signals retain enough class-discriminative information for a classifier trained on that same distribution.\n\nThe paper acknowledges this limitation in the Discussion ('for real-world scans we evaluate digitization indirectly through downstream diagnostic performance'), but the abstract and conclusions present the combined results as validating the full pipeline without clearly distinguishing which evidence supports which claim. The 95","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"The manuscript presents ECGLight, an end-to-end pipeline for digitizing paper ECG images and performing downstream myocardial infarction (MI) screening. The digitization pipeline uses YOLOv11x-seg for patched waveform segmentation, a YOLOv11x detector for lead-name and reference-pulse localization, and reference-pulse-based calibration to recover physically scaled 12-lead time-series. The digitizer is trained on synthetic paper ECG images rendered from PTB-XL waveforms using ECG-Image-Kit, and applied without retraining to real hospital paper ECGs from the ECG-Matrix dataset. Downstream classification benchmarks (MI vs. Normal on PTB-XL; Pre- vs. Post-procedural MI and OMI vs. non-OMI on ECG-Matrix) are performed using compact time-series models (e.g., Rocket, Arsenal, InceptionTime) under CPU-only inference, with SHAP-based interpretability analysis via an MLP surrogate. The system runs end-to-end in approximately 25–30 seconds per ECG on a laptop CPU.","tokens_in":36294,"tokens_out":1815,"duration_ms":465461,"significance":"The paper addresses a practically important problem: enabling automated ECG analysis in low-resource settings where paper printouts are still the norm and GPU/cloud infrastructure is unavailable. The end-to-end design connecting digitization, calibration, layout inference, classification, and interpretability within a CPU-only pipeline is a genuine engineering contribution. The provision of publicly available code, pretrained models, and a web dashboard is a notable strength that enhances reproducibility. The SHAP-based physiological attribution analysis, cross-validated by per-lead classification performance, adds interpretive value. The application to the ECG-Matrix dataset for OMI detection, a clinically relevant endpoint, demonstrates the pipeline's potential utility beyond benchmark datasets.","major_comments":[{"comment":"§V (Results, 'Overall ECG Digitization Quality') and Table I(d): The digitization quality is validated almost entirely on synthetic images (PTB-XL rendered via ECG-Image-Kit), with an 11.94% failure rate, Pearson r=0.806, and SNR=4.54 dB. The headline 95.51% MI classification accuracy on PTB-XL (Table III) is derived from signals that passed through this synthetic-to-synthetic digitization pipeline. The manuscript should explicitly state in the abstract and results that the 95.51% figure reflects classification on synthetic-to-synthetic digitized signals, not real paper ECG digitization. As written, the abstract's juxtaposition of '95.51% accuracy on PTB-XL' and '88.89% on ECG-Matrix' implies comparable evidential weight, which is misleading. This is a load-bearing framing issue for the central claim of real-world generalization.","section":null},{"comment":"§IV ('Benchmarking Digitized ECG-Based Time-Series Classification Models') and Table II(b): For ECG-Matrix (the only real-world test), the classifiers (e.g., Rocket for OMI detection) are trained directly on the digitized ECG-Matrix signals with patient-wise splits. Because both train and test signals come from the same digitization distribution, any systematic artifacts introduced by the synthetic-to-real domain gap would be present in both splits. The classifier could learn to exploit or tolerate these artifacts without them affecting accuracy. Thus, the 88.89% OMI accuracy does not, by itself, establish that the digitizer faithfully recovers true waveform morphology from real paper ECGs. The manuscript should acknowledge this limitation more prominently and discuss what specific validation (e.g., expert visual review of digitized ECG-Matrix signals, or signal-level comparison on a小型手动","section":null},{"comment":"§IV ('Data Generation') and §V ('YOLO-Based Patched Segmentation'): The synthetic-to-real domain gap is the central load-bearing assumption of the paper. The manuscript states that the digitizer 'generalized to real hospital-style ECG Matrix scans' (§VI), but provides no signal-level ground truth for ECG-Matrix to verify this. While the lack of paired ground truth is understandable, the paper could strengthen its claim by including qualitative examples of digitized real ECG-Matrix signals alongside the original scans, or by reporting any expert review of the digitized outputs. Without any such evidence, the claim of generalization rests entirely on indirect downstream classification accuracy, which is an imperfect proxy as noted above.","section":null},{"comment":"Table I(d): The 11.94% failure rate in digitization is reported but not reflected in the downstream classification metrics. It is unclear whether failed digitizations were excluded from the classification datasets (Table IIb), and if so, how failure was determined for real ECG-Matrix images where no ground truth exists. If failed digitizations were included, they may have introduced artifacts that affected classifier training and evaluation. The manuscript should clarify the handling of digitization failures in the downstream pipeline.","section":null}],"minor_comments":[{"comment":"Abstract: 'acute coronary occlusion (ACS) is overlooked' — ACS encompasses more than occlusion; consider rephrasing to 'acute coronary syndrome (ACS) is overlooked' or 'acute coronary occlusion is overlooked' for precision.","section":null},{"comment":"Table I(a): The YOLOv12x (Patched) row reports box precision 0.618 and recall 0.594, which are substantially lower than YOLOv11x. The text in §IV states 'YOLOv11x remained superior in precision/recall,' but the magnitude of YOLOv12x's underperformance is surprising. A brief discussion of why YOLOv12x performed so poorly on this task would be informative.","section":null},{"comment":"Table S2: The SNR comparison with challenge submissions is difficult to interpret because the proposed method's SNR (4.54 dB) is reported as a single value, while other methods report clean and deteriorated values. It is unclear whether 4.54 dB corresponds to clean, deteriorated, or an average. Clarification is needed for a fair comparison.","section":null},{"comment":"§V ('MLP-SHAP-Based Feature Analysis'): The MLP architecture is described as 'one hidden layer (100 neurons)' in the main text but as 'two fully connected hidden layers (100 and 50 neurons)' in the Supplementary Information (§'MLP-Based SHAP Feature Importance Analysis'). These should be reconciled.","section":null},{"comment":"Figure 2: The SHAP attribution maps are described as 'idealized 12-lead heartbeat' with overlaid SHAP traces. It would help to clarify whether the underlying heartbeat is a representative example or a schematic, and whether the SHAP values are from a single test beat or aggregated.","section":null},{"comment":"Table II(b): The 'Avg. Timesteps' column for full-sequence settings reports values like 1,633 and 4,046, which appear to be the number of time steps per recording. This should be labeled more clearly (e.g., 'Timesteps per record') to avoid confusion with the segmented setting's 280.","section":null},{"comment":"§VII (Conclusion): 'helps bring advanced diagnostic capabilities to settings where they in settings where digital ECG export...' — there is a duplicated phrase ('in settings where they in settings where').","section":null},{"comment":"References: Several references (e.g., [33] YOLO, [36] YOLO-Patch-Based-Inference) cite GitHub repositories or software without version numbers or specific commit hashes. For reproducibility, consider adding version or commit information.","section":null}],"recommendation":"major_revision","confidential_remarks":"The core engineering work is solid and the public code release is commendable. However, the evidential gap between the synthetic-to-synthetic PTB-XL results and the real-world ECG-Matrix results is the central vulnerability. The authors need to either (a) temper their claims about real-world generalization to match the evidence, or (b) provide additional validation (even qualitative) of digitization quality on real paper ECGs. The current framing, especially in the abstract, risks overstating what has been demonstrated. If the authors can address the framing and provide some signal-level evidence (even without ground truth, e.g., expert review or visual comparison), this could become a strong contribution. The paper is appropriate in scope for the journal."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive review. The referee's four major comments all concern the same underlying issue: the evidential asymmetry between our synthetic-to-synthetic PTB-XL results and our synthetic-to-real ECG-Matrix results, and whether the manuscript's framing adequately communicates this asymmetry. We agree with the substance of all four comments and will revise the manuscript accordingly. Specifically, we will: (1) clarify in the abstract and results that the 95.51% PTB-XL accuracy reflects synthetic-to-synthetic digitization, not real paper ECG digitization; (2) add a prominent limitation discussion acknowledging that ECG-Matrix classification accuracy does not, by itself, establish faithful waveform recovery from real paper ECGs, since systematic digitization artifacts would be present in both train and test splits; (3) add qualitative examples of digitized real ECG-Matrix signals alongside their original scans, and report the results of an expert visual review; and (4) clarify how digitization failures were handled in the downstream classification pipeline for both datasets. We do not have standing objections to any of the referee's comments.","responses":[{"response":"The referee is correct. The 95.51% accuracy on PTB-XL reflects classification on signals that were rendered from digital PTB-XL waveforms into synthetic paper images and then digitized back—the digitizer operates on synthetic images and is evaluated against the original digital ground truth. This is a synthetic-to-synthetic round-trip and does not test real paper ECG digitization. The 88.89% OMI accuracy on ECG-Matrix, by contrast, involves real hospital paper ECG scans, but as the referee notes in Comment 2, it has its own evidential limitations. We agree that the current abstract framing implies comparable evidential weight between these two numbers, which is misleading. We will revise the abstract to explicitly state that the PTB-XL result reflects classification on synthetic-to-synthetic digitized signals, and that the ECG-Matrix result reflects application to real paper ECGs without paired signal ground truth. We will also add a clarifying note in the Results section (§V) at the point where both numbers are first presented together.","revision_made":"yes","referee_comment":"§V and Table I(d): The 95.51% MI classification accuracy on PTB-XL is derived from synthetic-to-synthetic digitization, and the abstract's juxtaposition with 88.89% on ECG-Matrix implies comparable evidential weight, which is misleading."},{"response":"We fully agree with this analysis. Because both the training and test splits for ECG-Matrix classification pass through the same digitization pipeline, any systematic artifacts introduced by the synthetic-to-real domain gap would be present in both splits. The classifier could learn to exploit or tolerate these artifacts, meaning that downstream classification accuracy is an imperfect proxy for digitization fidelity. This is a genuine limitation of our validation strategy for ECG-Matrix, and it stems from the fundamental absence of paired digital waveform ground truth for real paper ECGs—a limitation we acknowledge in the Discussion (§VI) but do not sufficiently emphasize. We will add a prominent paragraph in the Discussion explicitly stating that ECG-Matrix classification accuracy does not, by itself, establish faithful waveform morphology recovery, and that the domain gap could introduce systematic artifacts that are invisible to the downstream classification metric. We will also discuss what additional validation would be needed (expert visual review, signal-level comparison on a manually digitized subset) and note that we are incorporating the first of these in the revision (see our response to Comment 3).","revision_made":"yes","referee_comment":"§IV and Table II(b): For ECG-Matrix, classifiers are trained on digitized signals from the same digitization distribution in both train and test splits, so systematic artifacts from the synthetic-to-real domain gap could be learned or tolerated without affecting accuracy. The 88.89% OMI accuracy does not establish faithful waveform recovery from real paper ECGs."},{"response":"This is a fair criticism. The claim that the digitizer generalized to real ECG-Matrix scans currently rests on indirect evidence (downstream classification accuracy and physiologically plausible SHAP attributions), which, as the referee correctly notes, is an imperfect proxy. We will address this in two ways. First, we will add a new supplementary figure showing qualitative examples of digitized real ECG-Matrix signals displayed alongside the original scanned paper ECG images, so readers can visually assess reconstruction quality. Second, we have conducted an expert visual review of a sample of digitized ECG-Matrix signals by a cardiologist on our team (F.G.), and we will report the results of this review in the revised manuscript, including the fraction of signals rated as clinically interpretable and the types of artifacts observed. We acknowledge that even expert review is not a substitute for signal-level ground truth, and we will state this explicitly. However, we believe these additions substantially strengthen the generalization claim beyond indirect downstream metrics alone.","revision_made":"yes","referee_comment":"§IV and §V: The synthetic-to-real domain gap is the central load-bearing assumption. The manuscript claims the digitizer 'generalized to real hospital-style ECG Matrix scans' but provides no signal-level ground truth or qualitative examples for ECG-Matrix."},{"response":"The referee raises an important point about the handling of digitization failures in the downstream pipeline. We will clarify this in the revised manuscript. For PTB-XL, the 11.94% failure rate reported in Table I(d) was measured on the 1,600-image digitization evaluation subset (which spans all six diagnostic superclasses), while the classification datasets (Table IIb) were constructed from separate Normal and MI subsets. Digitization failures on the classification subsets were identified by automated quality-control checks: signals with missing leads (fewer than 12 leads successfully extracted), implausible calibration values (reference pulse detection failure or out-of-range scaling factors), or signals with zero or near-zero variance were flagged as failures and excluded from the classification datasets. For ECG-Matrix, the same automated quality-control checks were applied, since no signal-level ground truth exists. We will add a paragraph in §IV (Data Generation / Dataset Construction) explicitly describing these quality-control criteria and reporting the fraction of images excluded from each classification dataset due to digitization failure. We will also add a note in Table IIb indicating the number of records excluded prior to the reported splits.","revision_made":"yes","referee_comment":"Table I(d): The 11.94% digitization failure rate is not reflected in downstream classification metrics. It is unclear whether failed digitizations were excluded from classification datasets, and if so, how failure was determined for real ECG-Matrix images where no ground truth exists."}],"tokens_in":36131,"tokens_out":1441,"duration_ms":218655,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"Here's what you need to know: this paper integrates YOLO-based ECG segmentation, reference-pulse calibration, layout inference, and downstream MI classification into a single CPU-only pipeline that runs in ~25–30 seconds. The code and pretrained models are public on GitHub. The headline 95.51% MI accuracy comes from a synthetic-to-synthetic loop (train and test on ECG-Image-Kit renderings of PTB-XL), while the only real-world test (ECG-Matrix, 88.89% OMI accuracy) lacks signal-level ground truth. The stress-test concern about this conflation is correct and lands squarely on the paper as written — the abstract presents both numbers as validating the full pipeline without distinguishing which evidence supports which claim. The Discussion section does acknowledge the limitation, but the framing in the abstract and conclusions overstates what was shown. That said, the paper does several things well. The patched YOLOv11x segmentation with multi-grid fusion is a reasonable engineering choice, and the mask-level IoU improvement over full-image inference (0.647 vs. 0.221) is meaningful. Reference-pulse-based calibration instead of grid-line detection is a sensible design decision for degraded scans. The layout inference across 16 formats including Cabrera is thorough. The SHAP attribution analysis is internally consistent — attributions concentrate on QRS and ST-T segments in clinically expected lead territories, and the per-lead classification cross-check (Table S4) corroborates the rankings. The broad classifier benchmark (9 models, three endpoints, patient-wise splits, latency reporting) is useful reference material. The soft spots are real but mostly proportional. The 11.94% digitization failure rate is mentioned but not carried into downstream metrics — we don't know whether failed digitizations were excluded from classification or how they'd affect results. The MLP architecture description is inconsistent between the main text (one hidden layer, 100 neurons) and the supplement (two layers, 100 and 50 neurons, with 50 PCA components for SHAP). The Table S2 comparison against PhysioNet Challenge entries is not under identical conditions and shouldn't be read as a head-to-head. The ECG-Matrix sample sizes are small (107 test patients for OMI), so the 88.89% has wide uncertainty bounds. None of these are fatal — they're the kind of gaps a good revision would close. This paper is for researchers and engineers working on ECG digitization pipelines, particularly those targeting low-resource deployment. It deserves a serious referee who can push the authors to (1) separate synthetic vs. real-world claims cleanly in the abstract, (2) report what happens to downstream metrics when digitization failures are included, and (3) reconcile the MLP description. The core engineering contribution is legitimate and the code is available, which raises the bar for reproducibility.","headline":"Solid engineering pipeline for paper ECG digitization with CPU-only deployment; headline accuracy conflates synthetic-to-synthetic and real-world evidence","tokens_in":37053,"tokens_out":673,"would_cite":false,"duration_ms":168916,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Paper ECGs to diagnosis in 30 seconds on a laptop","keywords":["ECG digitization","paper ECG","myocardial infarction","YOLOv11","CPU-only inference","edge AI","reference pulse calibration","SHAP interpretability"],"falsifier":"If a set of real hospital paper ECGs with paired digital ground-truth signals were digitized by this pipeline and showed substantially lower Pearson correlation or higher RMSE than the synthetic test set, the claim of cross-domain generalization would be weakened.","tokens_in":36053,"feed_emoji":"🫀","tokens_out":698,"duration_ms":107565,"temperature":0.7,"pith_summary":"The paper presents a complete software pipeline that takes a smartphone photo or scan of a paper ECG printout and, running entirely on a laptop CPU with no internet connection, converts it into a calibrated 12-lead digital signal and then screens for myocardial infarction. The digitization stage uses a YOLOv11-based instance segmentation model trained on synthetic paper-ECG images rendered from the PTB-XL waveform dataset. The key claim is that this model, having seen only synthetic images, generalizes to real hospital paper ECG scans from a different institution without retraining, producing signals accurate enough to train downstream MI classifiers. The full pipeline runs in 25 to 30 seconds per ECG on CPU-only hardware, achieving 95.51% accuracy for MI detection on PTB-XL and 88.89% accuracy for occlusion MI detection on the real-world ECG-Matrix dataset. The authors also use SHAP attributions on a simple MLP classifier to show that the model's predictions are driven by physiologically expected ECG regions, the QRS complex and ST segment, rather than digitization artifacts.","feed_headline":"Paper ECG to MI diagnosis in 30 seconds, CPU-only","feed_subtitle":"A YOLO-based pipeline trained on synthetic images generalizes to real hospital paper ECGs, running entirely on a laptop without internet or","key_machinery":"YOLOv11","core_discovery":"The central mechanism is a three-stage pipeline: (1) patch-based YOLOv11 instance segmentation to isolate inked waveforms from grid lines, shadows, and noise, with multi-scale patch fusion to preserve fine trace boundaries; (2) automated lead-name detection and layout inference across 16 standard ECG formats, including Cabrera ordering, to correctly map each trace to its anatomical lead; and (3) reference-pulse detection, a 1 mV, 200 ms calibration marker printed on most ECGs, to convert pixel coordinates into physically meaningful millivolts and milliseconds without relying on grid visibility. The discovery is that a digitizer trained exclusively on synthetic rendered images can generalize,","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Smartphone photo of paper ECG yields MI screening in 30 seconds","CPU-only pipeline turns paper ECG photos into 12-lead MI diagnosis","ECGLight digitizes paper ECGs and detects MI in under 30 seconds","Paper ECG photos to 12-lead signals for MI screening on CPU-only","Synthetic-trained pipeline digitizes paper ECGs for MI detection"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The central load-bearing assumption is that synthetic paper ECG images rendered from PTB-XL waveforms are realistic enough to train a digitizer that works on real hospital paper ECG scans. For the real-world ECG-Matrix dataset, there is no paired digital ground-truth signal, so digitization quality is only assessed indirectly through downstream classification accuracy, leaving open whether systematic digitization artifacts in real scans are inflating or distorting the MI-d","fun_headline_variants_meta":{"raw":{"variants":["Smartphone photo of paper ECG yields MI screening in 30 seconds","CPU-only pipeline turns paper ECG photos into 12-lead MI diagnosis","ECGLight digitizes paper ECGs and detects MI in under 30 seconds","Paper ECG photos to 12-lead signals for MI screening on CPU-only","Synthetic-trained pipeline digitizes paper ECGs for MI detection","Smartphone scan to calibrated 12-lead ECG for MI screening","Compute-light pipeline converts paper ECG to MI diagnosis in 30s","Legacy paper ECGs digitized for MI screening without internet","ECGLight: On-device digitization of paper ECGs for MI screening","Paper ECG to MI detection in 30 seconds using CPU-only resources"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":1641,"prompt_tokens":690,"completion_tokens":951,"prompt_tokens_details":null},"tokens_in":690,"tokens_out":951,"duration_ms":42356,"temperature":1.0,"reasoning_tokens":766,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T02:30:46.161948+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If a set of real hospital paper ECGs with paired digital ground-truth signals were digitized by this pipeline and showed substantially lower Pearson correlation or higher RMSE than the synthetic test set, the claim of cross-domain generalization would be weakened.","supporting_citations":[],"review_version":1}