Pith. sign in

REVIEW 3 major objections 6 minor 31 references

A single set of ECG features extracted from raw waveforms keeps prediction performance when models move between hospitals.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-30 22:56 UTC pith:E2QRSILG

load-bearing objection Solid two-cohort ECG transfer benchmark with clean pre-specified AUROC tests; the age/sex confound is real but does not sink the result, and the paper is worth engaging. the 3 major comments →

arxiv 2607.23412 v1 pith:E2QRSILG submitted 2026-07-26 cs.LG

Harmonized Interpretable ECG Waveform Features for Robust Cross-Dataset Clinical Prediction

Jie Lin , Weijie Sun , Sunil V. Kalmady , Anita Khalafbeigi , Abram Hindle , Padma Kaul , Russell Greiner This is my paper
classification cs.LG
keywords electrocardiogram (ECG)heart failurefeature harmonizationexternal validationdataset shiftinterpretable machine learningXGBoostcross-dataset generalization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Hospital ECG models often break when moved to a new site because devices and vendor measurement software do not define the same variables the same way. This paper builds one fixed, interpretable feature set directly from the raw 12-lead waveform—beat morphology, heart-rate variability, plus compact autoregressive and wavelet descriptors—and uses it on both MIMIC-IV and a large Alberta cohort. On heart-failure classification and two 30-day mortality tasks, XGBoost models trained on that shared space keep external AUROC within 10% of source-site internal AUROC in every transfer direction, and match the discrimination of each site’s own vendor measurements internally. An end-to-end waveform network scores higher in absolute terms, but the handcrafted interface remains competitive on relative transfer stability and stays clinically readable. The practical point is that a consistent waveform-derived feature layer makes fair external validation possible and reduces the need to re-engineer features at every hospital.

Core claim

When ECG predictors are computed from raw waveforms with identical definitions across sites, external AUROC retains at least 90% of source-site internal AUROC on heart failure and 30-day mortality tasks in both transfer directions between MIMIC-IV and Alberta, while internal AUROC stays within about 2.5% of each cohort’s vendor machine-measurement baselines.

What carries the argument

The harmonized FeatureDB+AR/WV representation: morphology/interval and HRV summaries from leads II and V5, plus low-order autoregressive coefficients and wavelet-variance features from leads II and V1, all extracted from fixed-length raw segments with one shared pipeline, then modeled with XGBoost under patient-disjoint internal and bidirectional external tests.

Load-bearing premise

The paper assumes that heart-failure labels, 30-day death ascertainment, sinus-rhythm flags, and quality exclusions are similar enough between the two hospital systems that retained AUROC mainly reflects the shared feature interface rather than leftover label or case-mix mismatch.

What would settle it

Train the same FeatureDB+AR/WV pipeline on a third independent ECG cohort with different devices and labeling rules; if external AUROC falls more than 10% relative to source internal AUROC on the same three tasks, the retention claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Cross-hospital ECG validation can use one fixed waveform-derived feature table instead of aligning incompatible vendor measurement dumps.
  • Sites can audit risk models with standard morphology, HRV, and time–frequency terms rather than opaque saliency maps alone.
  • AUROC can look stable under transfer while AUPRC still shifts with prevalence, so target-site recalibration remains necessary before deployment.
  • Handcrafted harmonized features are a practical baseline against which end-to-end waveform models should be judged for transfer, not only for peak internal score.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the interface is truly vendor-agnostic, open multi-hospital feature warehouses become feasible without waiting for manufacturers to standardize measurement APIs.
  • The same extraction recipe could be stress-tested on other outcomes (e.g., reduced ejection fraction or arrhythmia subtypes) to see whether retention generalizes beyond HF and short-term death.
  • Larger AUPRC swings under transfer suggest that threshold and calibration protocols, not just feature identity, will dominate real-world safety cases.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The manuscript studies whether a fixed tabular representation computed from raw ECG waveforms can provide more reliable cross-site prediction than vendor-specific machine measurements. FeatureDB morphology/HRV features from leads II and V5 and AR/wavelet descriptors from leads II and V1 are combined with age and sex and used to train XGBoost models for heart failure, 30-day mortality, and sinus-rhythm 30-day mortality in MIMIC-IV and an Alberta cohort. Using patient-disjoint internal tests and bidirectional external tests, the authors pre-specify H1 (external/internal AUROC at least 0.90) and H2 (internal AUROC within 10% of dataset-native vendor features). Reported internal AUROC is 0.793–0.820, external AUROC is 0.745–0.781, AUROC retention is 0.908–0.985, and internal differences from vendor-measurement baselines are approximately ±2.5%. An exploratory ConvNeXt model has higher absolute AUROC but a broadly similar relative transfer pattern.

Significance. If the reported performance is attributable to the harmonized waveform interface and the cohort labels are sufficiently comparable, the work would provide practically useful evidence that transparent, identically defined ECG features can support external validation across health systems. The study has several important strengths: fixed feature definitions, patient-level splitting, bidirectional external validation, matched development-set sizes, bootstrap confidence intervals, vendor-measurement baselines, and a raw-waveform deep-learning benchmark. The feature choices are also largely taken from prior pipelines rather than tuned on the transfer tests. Its current significance is nevertheless constrained by the absence of an ablation isolating the waveform features from age/sex and by limited demonstration that labels, rhythm calls, and exclusions are harmonized across the two cohorts.

major comments (3)
  1. [Section III-B; Table II; Figure 5] The central result does not isolate waveform harmonization from demographic transportability. Section III-B states that the 373 ECG-derived variables are augmented with age and sex, and Table II evaluates this combined model; Figure 5 identifies age as a dominant heart-failure contributor. A stable age/sex–outcome relationship could therefore preserve external AUROC even if the ECG morphology, HRV, AR, and wavelet features transfer poorly. Please report age/sex-only and waveform-only models for every task and transfer direction, ideally with feature-family ablations as well. If these are not added, the title, abstract, and H1 interpretation should explicitly describe a demographic-augmented interface rather than attributing retention to waveform harmonization itself.
  2. [Section IV-D; Table II; Figure 4] H1 is assessed using only the point estimate of R_AUROC, whereas the 95% CIs in Table II are for the numerator and denominator separately. The closest result is heart failure MIMIC→Alberta, with retention 0.7446/0.8200=0.9080—only 0.008 above the pre-specified threshold. Separate AUROC intervals do not quantify uncertainty in their ratio. Please bootstrap the internal and external test predictions to obtain a CI for R_AUROC (or relative AUROC loss), and state whether support requires the point estimate or, more stringently, the lower confidence bound to exceed 0.90. If the intended analysis is descriptive, the wording “H1 is supported” should be qualified accordingly.
  3. [Sections III–IV; Figure 2; Table I] Cross-dataset retention is interpreted as evidence for feature-interface robustness, but the manuscript does not yet establish comparable labels and selection mechanisms. The exact ICD code set and ECG–episode linkage rule for heart failure, mortality ascertainment and right-censoring procedures in each jurisdiction, the source of sinus-rhythm labels, and per-stage quality/feature-extraction exclusion rates are not fully specified. Residual concordance in labels, case mix, or exclusions could contribute to the observed transfer. Please provide the exact task definitions and exclusion counts by cohort/task, compare selected populations, and add a sensitivity analysis on the most strictly harmonized subset feasible.
minor comments (6)
  1. [Section III-B and Section V] The claim that the representation is fully reproducible would be strengthened by releasing the exact feature dictionary for all 373 variables, extraction/configuration code, package versions, and split identifiers where governance permits. The cited FeatureDB repository does not by itself specify the complete adapted pipeline used here.
  2. [Section IV-B] The Alberta development set is randomly subsampled once to match MIMIC. Please state the random seed and preferably repeat the internal/external experiments over several label-stratified subsampling draws, reporting the range or mean performance to show that the conclusions are not seed-dependent.
  3. [Section V-A] Please state explicitly that winsorization cutoffs and imputation means are estimated only from the source training data and then applied unchanged to validation, internal-test, and external-test records. This is important for ruling out target-site information leakage.
  4. [Section V-C and Table II] The ConvNeXt benchmark uses all 12 leads and 49,152 waveform values, whereas the harmonized representation uses selected features from leads II, V1, and V5. Since the comparison therefore differs in lead coverage and input duration as well as representation, this limitation should be stated when describing the benchmark; a matched-lead ConvNeXt ablation would be informative but is not essential.
  5. [Section VII-A and Figure 5] SHAP values are computed on a stratified subsample containing 25% positives, unlike the natural test prevalence. Because SHAP magnitudes and rankings can depend on prevalence and feature correlation, please clarify this limitation and consider a natural-prevalence sensitivity analysis.
  6. [Tables I–II and references] Table I has formatting problems around the MIMIC entries and the promised bold effective development sizes are not visible in the supplied text. There is a typographical error in “competetive” in Section VI, spacing is missing in several headings, and references [1] and [5] appear to duplicate the same npj Digital Medicine article.

Circularity Check

0 steps flagged

No significant circularity: empirical held-out AUROC/AUPRC benchmarks with a priori features and pre-specified inequality hypotheses.

full rationale

This is a standard supervised-learning transfer study. The load-bearing claims (H1: external AUROC retains ≥90% of source internal AUROC; H2: harmonized internal AUROC within 10% of machine-measurement baselines) are inequality checks on held-out, patient-disjoint internal and bidirectional external test sets, not identities forced by construction or by fitted parameters renamed as predictions. Feature definitions (FeatureDB morphology/HRV from Leads II/V5; AR/WV from II/V1) are fixed a priori from external pipelines and applied identically across cohorts; no data-driven selection on the test sets is reported. XGBoost hyperparameters are tuned on validation splits only. Self-citations to prior Alberta/MIMIC ECG work supply cohorts and task motivation but do not supply a uniqueness theorem or definitional premise that forces the retention results. Age/sex inclusion weakens causal attribution to waveform harmonization alone but is a validity issue, not circularity. The derivation chain is self-contained empirical evaluation against external benchmarks.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

Load-bearing content is almost entirely empirical protocol plus standard ML/ECG assumptions. No new physical entities. Free choices are classical featurization and training knobs (leads, AR/wavelet orders, winsorization, boosting hyperparameters). Domain assumptions include label linkage quality, R-peak/template reliability, and cross-cohort outcome comparability.

free parameters (5)
  • AR model order p=32 and retained coefficients N_AR=8 = p=32, N_AR=8
    Chosen following Nahak et al. configuration rather than derived; directly shapes the time-frequency half of the unified vector.
  • Stationary wavelet depth J=4 (Daubechies-2) and lead set {II,V1}/{II,V5} = J=4; FeatureDB on II+V5; AR/WV on II+V1
    A priori lead and wavelet settings fix which morphology/HRV/time-frequency content enters every model and transfer test.
  • 99% winsorization and training-mean imputation = 99% winsorize; mean / nan-category imputation
    Hand-chosen preprocessing that affects outlier and missingness handling across cohorts with different quality profiles.
  • XGBoost hyperparameters via 30-trial Optuna per task/dataset = selected per task/site on validation AUROC
    Fitted to validation AUROC (learning rate, depth, subsampling, regularization, etc.); necessary for reported performance though not part of H1’s definition.
  • H1/H2 thresholds (90% AUROC retention; 10% internal competitiveness) = R_AUROC≥0.90; relative internal gap≤10%
    Pre-specified success cutoffs that determine whether hypotheses are declared supported; not estimated from a utility model.
axioms (5)
  • domain assumption Identically defined waveform-derived tabular features are a valid shared interface for comparing models across ECG vendors and health systems.
    Core premise of Sections I and III; motivates discarding vendor measurements for cross-dataset work.
  • domain assumption ICD-based episode HF labels and 30-day all-cause mortality labels are sufficiently accurate and similarly targeted across MIMIC-IV and Alberta for discrimination comparisons.
    Task definitions in Section IV; no cross-cohort label adjudication study is provided.
  • domain assumption Rule-based R-peak detection plus average-beat templates yield stable morphology/HRV descriptors after the stated denoising/normalization.
    FeatureDB pipeline in Section III-B; failed detections are excluded rather than modeled.
  • domain assumption AUROC retention and AUPRC/AUROC with bootstrap CIs are adequate primary evidence for the pre-specified transfer hypotheses without clinical utility weights.
    Stated explicitly in Section IV-D evaluation metrics.
  • standard math Standard supervised learning and bootstrap uncertainty quantification apply to patient-disjoint ECG splits treated as i.i.d. within site after splitting.
    Used throughout modeling and CI reporting (Sections IV–VI).

pith-pipeline@v1.2.0-grok45-kimik3 · 17927 in / 3456 out tokens · 71680 ms · 2026-07-30T22:56:41.127046+00:00 · methodology

0 comments
read the original abstract

Electrocardiograms (ECGs) are widely used for cardiovascular risk prediction, yet models often fail to transfer across hospitals because of protocol, population, and measurement differences. We benchmark cross-dataset generalization on three tasks - heart failure classification, 30-day all-cause mortality, and 30-day mortality among sinus-rhythm ECGs - using two large cohorts (MIMIC-IV and the Alberta Cohort). To reduce vendor-specific measurement mismatch, we build a harmonized, interpretable feature representation computed directly from raw waveforms: FeatureDB morphology/heart-rate-variability summaries plus compact time-frequency descriptors (autoregressive and wavelet features). We train XGBoost models on this unified feature space and evaluate with patient-disjoint internal and bidirectional external testing. We pre-specify two hypotheses: (H1) external AUROC retains at least 90% of source-site internal AUROC under transfer, and (H2) internal AUROC of the harmonized feature set stays within 10% of dataset-native machine-measurement models. Across tasks, internal AUROC is 0.79-0.82 and cross-dataset AUROC is 0.74-0.78, with larger and direction-dependent AUPRC shifts under transfer. As an exploratory benchmark, an end-to-end ConvNeXt model trained directly on raw ECG waveforms with age and sex achieves higher internal AUROC, while the harmonized representation remains competitive in relative cross-dataset transfer stability. These findings show that a consistent waveform-derived feature interface preserves performance, supports realistic external validation, and provides a transparent alternative for cross-site clinical prediction.

Figures

Figures reproduced from arXiv: 2607.23412 by Abram Hindle, Anita Khalafbeigi, Jie Lin, Padma Kaul, Russell Greiner, Sunil V. Kalmady, Weijie Sun.

Figure 1
Figure 1. Figure 1: Overview of the study design, including the two ECG cohorts, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: ECG filtering and dataset splitting (MIMIC-IV and Alberta Cohort). [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Source-site internal AUROC comparison between the harmonized FeatureDB+AR/WV representation and dataset-native machine-derived ECG [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Relative external AUROC change from source-site internal performance for the main harmonized FeatureDB+AR/WV model and the end-to-end ECG [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: SHAP summary plots for Task 1 (heart failure classification) across internal evaluation and cross-dataset transfer settings. Only the top 10 features [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 1 linked inside Pith

  1. [1]

    Development and Validation of Machine Learning Algorithms Based on Electrocardiograms for Cardiovascular Diagnoses at the Pop- ulation Level,

    S. V . Kalmady, A. Salimi, W. Sun, N. Sepehrvand, Y . Nademi, K. Bainey et al., “Development and Validation of Machine Learning Algorithms Based on Electrocardiograms for Cardiovascular Diagnoses at the Pop- ulation Level,”npj Digital Medicine, vol. 7, no. 1, p. 133, 2024

    2024

  2. [2]

    Towards Artificial Intelligence-Based Learning Health System for Population-Level Mortality Prediction Using Electrocardiograms,

    W. Sun, S. V . Kalmady, N. Sepehrvand, A. Salimi, Y . Nademi, K. Bainey et al., “Towards Artificial Intelligence-Based Learning Health System for Population-Level Mortality Prediction Using Electrocardiograms,” npj Digital Medicine, vol. 6, no. 1, p. 21, 2023

    2023

  3. [3]

    Empirical investigation of multi-source cross-validation in clini- cal ECG classification,

    T. Leinonen, D. Wong, A. Vasankari, A. Wahab, R. Nadarajah, M. Kaisti et al., “Empirical investigation of multi-source cross-validation in clini- cal ECG classification,”Computers in Biology and Medicine, vol. 183, p. 109271, 2024

    2024

  4. [4]

    Automatic Diagnosis of the 12-Lead ECG Using a Deep Neural Network,

    A. H. Ribeiro, M. H. Ribeiro, G. M. M. Paix ˜ao, D. M. Oliveira, P. R. Gomes, J. A. Canazartet al., “Automatic Diagnosis of the 12-Lead ECG Using a Deep Neural Network,”Nature Communications, vol. 11, no. 1, p. 1760, 2020

    2020

  5. [5]

    Development and validation of machine learning algorithms based on electrocardiograms for cardiovascular diagnoses at the pop- ulation level,

    S. V . Kalmady, A. Salimi, W. Sun, N. Sepehrvand, Y . Nademi, K. Bainey et al., “Development and validation of machine learning algorithms based on electrocardiograms for cardiovascular diagnoses at the pop- ulation level,”NPJ Digital Medicine, vol. 7, no. 1, p. 133, 2024

    2024

  6. [6]

    ENCASE: An Ensemble Classifier for ECG Classification Using Expert Features and Deep Neural Networks,

    S. Hong, M. Wu, Y . Zhou, Q. Wang, J. Shang, H. Liet al., “ENCASE: An Ensemble Classifier for ECG Classification Using Expert Features and Deep Neural Networks,” inComputing in Cardiology (CinC), 2017

    2017

  7. [7]

    Combining Deep Neural Networks and Engineered Features for Cardiac Arrhythmia Detection from ECG Recordings,

    S. Hong, Y . Zhou, M. Wu, J. Shang, Q. Wang, H. Liet al., “Combining Deep Neural Networks and Engineered Features for Cardiac Arrhythmia Detection from ECG Recordings,”Physiological Measurement, vol. 40, no. 5, p. 054009, 2019

    2019

  8. [8]

    Detecting Long QT Syndrome and First-Degree Atrioventricular Block Using Single-Lead AI-ECG: A Multi-Center Real-World Study,

    S. Fan, D. Zhang, Y . Wang, S. Geng, K. Lu, M. Sanget al., “Detecting Long QT Syndrome and First-Degree Atrioventricular Block Using Single-Lead AI-ECG: A Multi-Center Real-World Study,”arXiv preprint arXiv:2502.17499, 2025

    arXiv 2025

  9. [9]

    Evaluation of Handcrafted Features and Learned Representations for the Classification of Arrhythmia and Congestive Heart Failure in ECG,

    S. Nahak, A. Pathak, and G. Saha, “Evaluation of Handcrafted Features and Learned Representations for the Classification of Arrhythmia and Congestive Heart Failure in ECG,”Biomedical Signal Processing and Control, vol. 79, p. 104230, 2023

    2023

  10. [10]

    Cardiologist-Level Arrhythmia Detection and Classification in Ambulatory Electrocardiograms Using a Deep Neural Network,

    A. Y . Hannun, P. Rajpurkar, M. Haghpanahi, G. H. Tison, C. Bourn, M. P. Turakhiaet al., “Cardiologist-Level Arrhythmia Detection and Classification in Ambulatory Electrocardiograms Using a Deep Neural Network,”Nature Medicine, vol. 25, no. 1, pp. 65–69, 2019

    2019

  11. [11]

    Prospective validation of a deep learning electrocardio- gram algorithm for the detection of left ventricular systolic dysfunction,

    Z. I. Attia, S. Kapa, X. Yao, F. Lopez-Jimenez, T. L. Mohan, P. A. Pellikkaet al., “Prospective validation of a deep learning electrocardio- gram algorithm for the detection of left ventricular systolic dysfunction,” Journal of Cardiovascular Electrophysiology, vol. 30, no. 5, pp. 668– 674, 2019

    2019

  12. [12]

    Opportunities and challenges of deep learning methods for electrocardiogram data: A systematic review,

    S. Hong, Y . Zhou, J. Shang, C. Xiao, and J. Sun, “Opportunities and challenges of deep learning methods for electrocardiogram data: A systematic review,”Computers in Biology and Medicine, vol. 122, p. 103801, 2020

    2020

  13. [13]

    The Clinician and Dataset Shift in Artificial Intelligence,

    S. G. Finlayson, A. Subbaswamy, K. Singh, J. Bowers, A. Kupke, J. Zit- trainet al., “The Clinician and Dataset Shift in Artificial Intelligence,” New England Journal of Medicine, vol. 385, no. 3, pp. 283–286, 2021

    2021

  14. [14]

    Generalization challenges in electrocardiogram deep learning: insights from dataset characteristics and attention mech- anism,

    Z. Huang, S. MacLachlan, L. Yu, L. F. Herbozo Contreras, N. D. Truong, A. H. Ribeiroet al., “Generalization challenges in electrocardiogram deep learning: insights from dataset characteristics and attention mech- anism,”Future Cardiology, vol. 20, no. 4, pp. 209–220, 2024

    2024

  15. [15]

    Heart rate variability: standards of measurement, physiological interpretation and clinical use,

    Task Force of the European Society of Cardiology and the North Amer- ican Society of Pacing and Electrophysiology, “Heart rate variability: standards of measurement, physiological interpretation and clinical use,” Circulation, vol. 93, no. 5, pp. 1043–1065, 1996

    1996

  16. [16]

    Recommendations for the Standardization and Inter- pretation of the Electrocardiogram,

    P. Kligfield, L. S. Gettes, J. J. Bailey, R. Childers, B. J. Deal, E. W. Hancocket al., “Recommendations for the Standardization and Inter- pretation of the Electrocardiogram,”Circulation, vol. 115, no. 10, pp. 1306–1324, 2007

    2007

  17. [17]

    AF classification from a short single lead ECG recording: The PhysioNet/Computing in Cardiology Challenge 2017,

    G. D. Clifford, C. Liu, B. Moody, H.-L. Li-wei, I. Silva, Q. Li et al., “AF classification from a short single lead ECG recording: The PhysioNet/Computing in Cardiology Challenge 2017,” inProceedings of the 2017 Computing in Cardiology (CinC), 2017, pp. 1–4

    2017

  18. [18]

    Cardiac arrhythmia classifi- cation using autoregressive modeling,

    D. Ge, N. Srinivasan, and S. M. Krishnan, “Cardiac arrhythmia classifi- cation using autoregressive modeling,”BioMedical Engineering OnLine, vol. 1, p. 5, 2002

    2002

  19. [19]

    Combining low- dimensional wavelet features and support vector machine for arrhythmia beat classification,

    Q. Qin, J. Li, L. Zhang, Y . Yue, and C. Liu, “Combining low- dimensional wavelet features and support vector machine for arrhythmia beat classification,”Scientific Reports, vol. 7, no. 1, p. 6067, 2017

    2017

  20. [20]

    Calibration: the Achilles heel of predictive analytics,

    B. Van Calster, D. J. McLernon, M. van Smeden, L. Wynants, E. W. Steyerberg, P. Bossuytet al., “Calibration: the Achilles heel of predictive analytics,”BMC Medicine, vol. 17, no. 1, p. 230, 2019

    2019

  21. [21]

    Changing predictor measurement procedures affected the performance of prediction models in clinical examples,

    K. Luijken, L. Wynants, M. van Smeden, B. Van Calster, E. W. Steyerberg, R. H. Groenwoldet al., “Changing predictor measurement procedures affected the performance of prediction models in clinical examples,”Journal of Clinical Epidemiology, vol. 119, pp. 7–18, 2020

    2020

  22. [22]

    Comparison of automated interval measurements by widely used algorithms in digital electrocardiographs,

    P. Kligfield, F. Badilini, I. Denjoy, S. Babaeizadeh, E. Clark, J. De Bie et al., “Comparison of automated interval measurements by widely used algorithms in digital electrocardiographs,”American Heart Journal, vol. 200, pp. 1–10, 2018

    2018

  23. [23]

    Research on cross- dataset cardiac signal domain generalization and feature interpretability,

    R. Li, Y . Aierken, Y . Xu, J. Liu, and Y . Tang, “Research on cross- dataset cardiac signal domain generalization and feature interpretability,” Scientific Reports, vol. 16, no. 1, p. 3138, 2025

    2025

  24. [24]

    MIMIC-IV-ECG: Diagnostic Electrocardiogram Matched Subset,

    B. Gow, T. Pollard, L. A. Nathanson, A. Johnson, B. Moody, C. Fer- nandeset al., “MIMIC-IV-ECG: Diagnostic Electrocardiogram Matched Subset,” 2023, physioNet, Version 1.0. [25]Philips DXL ECG Algorithm: Physician’s Guide, Revision e ed., Philips Medical Systems, Nov. 2016

    2023

  25. [26]

    Evaluation and management of athletes with long QT syndrome,

    A. T. Gomez, J. M. Prutkin, and A. L. Rao, “Evaluation and management of athletes with long QT syndrome,”Sports Health, vol. 8, no. 6, pp. 527–535, 2016

    2016

  26. [27]

    XGBoost: A Scalable Tree Boosting System,

    T. Chen and C. Guestrin, “XGBoost: A Scalable Tree Boosting System,” inProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016, pp. 785–794

    2016

  27. [28]

    R. R. Wilcox,Introduction to Robust Estimation and Hypothesis Testing, 4th ed. Academic Press, 2017

    2017

  28. [29]

    R. J. A. Little and D. B. Rubin,Statistical Analysis with Missing Data, 3rd ed. Wiley, 2019

    2019

  29. [30]

    A Unified Approach to Interpreting Model Predictions,

    S. M. Lundberg and S.-I. Lee, “A Unified Approach to Interpreting Model Predictions,” inProceedings of the 31st International Conference on Neural Information Processing Systems, 2017, pp. 4768–4777

    2017

  30. [31]

    Optuna: A Next-Generation Hyperparameter Optimization Framework,

    T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama, “Optuna: A Next-Generation Hyperparameter Optimization Framework,” inPro- ceedings of the 25th ACM SIGKDD International Conference on Knowl- edge Discovery and Data Mining, 2019, pp. 2623–2631

    2019

  31. [32]

    Convnext v2: Co-designing and scaling convnets with masked autoen- coders,

    S. Woo, S. Debnath, R. Hu, X. Chen, Z. Liu, I. S. Kweonet al., “Convnext v2: Co-designing and scaling convnets with masked autoen- coders,”arXiv preprint arXiv:2301.00808, 2023. 39th IEEE International Symposium on Computer-Based Medical Systems (CBMS 2026)

    Pith/arXiv arXiv 2023