pith. sign in

arxiv: 2604.09782 · v1 · submitted 2026-04-10 · 💻 cs.CV

Biomarker-Based Pretraining for Chagas Disease Screening in Electrocardiograms

Elias Stenhede , Arian Ranjbar This is my paper

Pith reviewed 2026-05-10 16:45 UTC · model grok-4.3

classification 💻 cs.CV
keywords Chagas diseaseelectrocardiogrampretrainingblood biomarkerstransfer learningECG classificationPhysioNet challenge
0
0 comments X

The pith

Pretraining an ECG model to predict blood biomarkers from one dataset transfers to improved Chagas disease detection in another.

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

The paper addresses the challenge of limited and noisy labels for Chagas disease screening using electrocardiograms by proposing a two-stage training process. An ECG feature extractor is first trained on the large MIMIC-IV-ECG dataset to predict percentile-binned blood biomarkers. This pretrained model is then fine-tuned on Brazilian ECG datasets to classify Chagas disease. The resulting five-model ensemble reached a score of 0.269 on the hidden test set, placing fifth in the 2025 PhysioNet Challenge. The work shows how auxiliary prediction tasks on abundant ECG data can support performance on rare-disease tasks with scarce labels.

Core claim

The authors establish that an ECG feature extractor pretrained to predict blood biomarkers from the MIMIC-IV-ECG dataset can be successfully fine-tuned on Brazilian datasets to detect Chagas disease, yielding competitive performance in the PhysioNet Challenge.

What carries the argument

biomarker-based pretraining, in which the model learns ECG representations by predicting percentile-binned blood biomarkers before fine-tuning for Chagas classification

Where Pith is reading between the lines

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

  • The same pretraining step could support detection of other cardiac conditions where labeled ECG data is scarce.
  • Applying the approach to additional populations would test whether the learned biomarker-ECG links generalize beyond the original datasets.
  • Combining this supervised pretraining with other self-supervised ECG methods might produce stronger starting representations.

Load-bearing premise

The relationships between ECG signals and blood biomarkers observed in the MIMIC-IV dataset also hold sufficiently in the Brazilian Chagas patient data to allow useful feature transfer.

What would settle it

Training identical models from random initialization on the Brazilian Chagas datasets and comparing detection performance to the biomarker-pretrained versions; consistent lack of improvement from pretraining would refute the transfer benefit.

Figures

Figures reproduced from arXiv: 2604.09782 by Arian Ranjbar, Elias Stenhede.

Figure 1
Figure 1. Figure 1: Shows the predicted probability distributions over blood test percentiles for four ECGs taken from four different [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
read the original abstract

Chagas disease screening via ECGs is limited by scarce and noisy labels in existing datasets. We propose a biomarker-based pretraining approach, where an ECG feature extractor is first trained to predict percentile-binned blood biomarkers from the MIMIC-IV-ECG dataset. The pretrained model is then fine-tuned on Brazilian datasets for Chagas detection. Our 5-model ensemble, developed by the Ahus AIM team, achieved a challenge score of 0.269 on the hidden test set, ranking 5th in Detection of Chagas Disease from the ECG: The George B. Moody PhysioNet Challenge 2025. Source code and the model are shared on GitHub: github.com/Ahus-AIM/physionet-challenge-2025

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes a biomarker-based pretraining approach for ECG-based Chagas disease screening. An ECG feature extractor is first trained on the MIMIC-IV-ECG dataset to predict percentile-binned blood biomarkers, then fine-tuned on Brazilian Chagas datasets. A 5-model ensemble achieves a challenge score of 0.269 on the hidden test set, ranking 5th in the George B. Moody PhysioNet Challenge 2025. Source code is shared on GitHub.

Significance. If the pretraining step is shown to drive the performance, the work demonstrates a practical route to mitigate label scarcity for Chagas cardiomyopathy detection by transferring representations learned from abundant biomarker-annotated ICU ECGs. The external validation via the challenge ranking provides a concrete empirical anchor. However, the absence of controls leaves open whether the biomarker objective adds value beyond generic ECG pretraining.

major comments (3)
  1. [Methods] Methods: The manuscript provides no ablation comparing the biomarker-pretrained model against a from-scratch baseline or alternative pretraining objectives (e.g., contrastive or reconstruction) on the Chagas fine-tuning task. Without this isolation, the central claim that biomarker pretraining improves Chagas detection cannot be substantiated.
  2. [Results] Results: No statistical tests, confidence intervals, or variance estimates are reported for the ensemble score of 0.269, nor details on model selection or ensembling procedure. This weakens the reliability of the 5th-place ranking as evidence for the method.
  3. [Methods] Methods: Architecture details, loss functions for the biomarker regression task, binning thresholds, and training hyperparameters are not fully specified, preventing assessment of whether the pretraining objective is well-posed or reproducible.
minor comments (2)
  1. [Abstract] Abstract and Methods: The specific blood biomarkers used for pretraining and the exact percentile binning procedure are not enumerated, which would aid interpretability of the learned features.
  2. [Introduction] Introduction: Limited discussion of prior transfer-learning or multi-task ECG pretraining literature; adding 2-3 key references would better situate the biomarker approach.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback, which highlights important areas for strengthening the evidence and reproducibility of our biomarker-based pretraining approach. We address each major comment below and have made revisions to the manuscript where feasible, including additional experiments and expanded details.

read point-by-point responses

  1. Referee: [Methods] Methods: The manuscript provides no ablation comparing the biomarker-pretrained model against a from-scratch baseline or alternative pretraining objectives (e.g., contrastive or reconstruction) on the Chagas fine-tuning task. Without this isolation, the central claim that biomarker pretraining improves Chagas detection cannot be substantiated.

    Authors: We agree that ablations are necessary to substantiate the specific benefit of the biomarker pretraining objective over generic alternatives. The original submission prioritized a competitive entry in the time-constrained PhysioNet Challenge. In the revision, we have added results from from-scratch baselines trained on the same Chagas fine-tuning data and report performance differences. We also include a brief discussion of why contrastive or reconstruction objectives were not used, noting that the supervised biomarker signal provides direct physiological alignment not guaranteed by self-supervised methods. These additions appear in the revised Methods and Results sections. revision: yes

  2. Referee: [Results] Results: No statistical tests, confidence intervals, or variance estimates are reported for the ensemble score of 0.269, nor details on model selection or ensembling procedure. This weakens the reliability of the 5th-place ranking as evidence for the method.

    Authors: We acknowledge that the original manuscript lacked uncertainty quantification and procedural details. The 0.269 score is the fixed official challenge metric on the hidden test set and cannot be altered. However, we have revised the Results section to describe the ensembling procedure (prediction averaging across five models with varied random seeds and stratified data splits) and added bootstrap-derived 95% confidence intervals computed on the internal validation set. We also include a note on model selection criteria. These changes provide better context for interpreting the ranking without overstating its statistical robustness. revision: yes

  3. Referee: [Methods] Methods: Architecture details, loss functions for the biomarker regression task, binning thresholds, and training hyperparameters are not fully specified, preventing assessment of whether the pretraining objective is well-posed or reproducible.

    Authors: We apologize for the incomplete specification. The revised Methods section now provides full details: the ECG encoder follows a ResNet-18 architecture adapted for 12-lead signals with specified kernel sizes and channel dimensions; the pretraining loss is mean squared error applied to the binned biomarker targets; binning uses dataset-specific percentiles (25th, 50th, 75th) for each of the 10 biomarkers; and all hyperparameters (learning rate schedule, batch size, epochs, optimizer, and data augmentation) are enumerated in a new supplementary table. The publicly shared GitHub code already implements these choices, and the text now matches the code for full reproducibility. revision: yes

Circularity Check

0 steps flagged

No circularity detected in pretraining-fine-tuning workflow

full rationale

The manuscript describes a standard two-stage pipeline: pretraining an ECG feature extractor on the independent MIMIC-IV-ECG dataset to regress percentile-binned blood biomarkers, followed by fine-tuning the resulting weights on separate Brazilian Chagas-labeled ECG datasets. No equations, fitted parameters, or self-citations are presented that reduce the final Chagas detection score to a quantity defined by the target task itself. The workflow relies on cross-dataset transfer without self-referential definitions or load-bearing citations to prior author work that would close a loop.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the transferability of ECG features learned from biomarker prediction; with only the abstract available, the ledger reflects the high-level assumptions stated in the method description.

free parameters (1)
  • percentile binning thresholds for biomarkers
    Discretization boundaries chosen to convert continuous biomarker values into prediction targets for pretraining.
axioms (1)
  • domain assumption ECG signals contain information predictive of blood biomarker levels
    Invoked to justify pretraining utility on MIMIC-IV-ECG before transfer to Chagas labels.

pith-pipeline@v0.9.0 · 5422 in / 1422 out tokens · 89538 ms · 2026-05-10T16:45:13.817166+00:00 · methodology

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

13 extracted references · 13 canonical work pages · 1 internal anchor

  1. [1]

    PhysioBank, PhysioToolkit, and PhysioNet: Components of a New Research Resource for Complex Physiologic Signals

    Goldberger AL, Amaral LAN, Glass L, Hausdorff JM, Ivanov PC, Mark RG, Mietus JE, Moody GB, Peng CK, Stanley HE. PhysioBank, PhysioToolkit, and PhysioNet: Components of a New Research Resource for Complex Physiologic Signals. Circulation June 2000;101(23). ISSN 0009-7322, 1524-4539

    work page 2000

  2. [2]

    Detection of Chagas Disease from the ECG: The George B

    Reyna MA, Koscova Z, Pavlus J, Weigle J, Saghafi S, Gomes P, Elola A, Hassannia MS, Campbell K, Bahrami Rad A, Ribeiro AH, Ribeiro AL, Sameni R, Clif- ford GD. Detection of Chagas Disease from the ECG: The George B. Moody PhysioNet Challenge 2025. In Comput- ing in Cardiology 2025, volume 52. 2025; 1–4

    work page 2025

  3. [3]

    Detection of Chagas Disease from the ECG: The George B

    Reyna MA, Koscova Z, Pavlus J, Saghafi S, Weigle J, Elola A, Seyedi S, Campbell K, Li Q, Bahrami Rad A, Ribeiro A, Ribeiro ALP, Sameni R, Clifford GD. Detection of Chagas Disease from the ECG: The George B. Moody PhysioNet Challenge 2025, 2025. ArXiv.2510.02202

    work page arXiv 2025

  4. [4]

    The epidemiology of Chagas disease in the Americas

    Cucunub ´a ZM, Guti ´errez-Romero SA, Ram ´ırez JD, Vel´asquez-Ortiz N, Ceccarelli S, Parra-Henao G, Henao- Mart´ınez AF, Rabinovich J, Bas ´a˜nez MG, Nouvellet P, Abad-Franch F. The epidemiology of Chagas disease in the Americas. The Lancet Regional Health Americas Septem- ber 2024;37:100881. ISSN 2667193X

    work page 2024

  5. [5]

    Longitudinal study of patients with chronic Chagas cardiomyopathy in Brazil (SaMi-Trop project): a cohort profile

    Cardoso CS, Sabino EC, Oliveira CDL, De Oliveira LC, Ferreira AM, Cunha-Neto E, Bierrenbach AL, Ferreira JE, Haikal DS, Reingold AL, Ribeiro ALP. Longitudinal study of patients with chronic Chagas cardiomyopathy in Brazil (SaMi-Trop project): a cohort profile. BMJ Open May 2016;6(5):e011181. ISSN 2044-6055, 2044-6055

    work page 2016

  6. [6]

    W, Sch¨on TB, Ribeiro ALP

    Ribeiro AH, Ribeiro MH, Paix ˜ao GMM, Oliveira DM, Gomes PR, Canazart JA, Ferreira MPS, Andersson CR, Macfarlane PW, Meira Jr. W, Sch¨on TB, Ribeiro ALP. Au- tomatic diagnosis of the 12-lead ECG using a deep neural network. Nature Communications April 2020;11(1):1760. ISSN 2041-1723

    work page 2020

  7. [7]

    MIMIC-IV-ECG: Diagnostic electrocar- diogram matched subset

    Gow B, Pollard T, Nathanson LA, Johnson A, Moody B, Fernandes C, Greenbaum N, Waks JW, Eslami P, Carbon- ati T, Chaudhari A, Herbst E, Moukheiber D, Berkowitz S, Mark R, Horng S. MIMIC-IV-ECG: Diagnostic electrocar- diogram matched subset

    work page

  8. [8]

    MIMIC-IV

    Johnson A, Bulgarelli L, Pollard T, Horng S, Celi LA, Mark R. MIMIC-IV

    work page

  9. [9]

    Muon: An optimizer for hidden layers in neural networks| Keller Jordan blog

    work page

  10. [10]

    Adam: A Method for Stochastic Optimization

    Kingma DP, Ba J. Adam: A method for stochastic opti- mization, January 2017. ArXiv:1412.6980

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  11. [11]

    InceptionTime: Finding AlexNet for time series classifica- tion, December 2020

    Fawaz HI, Lucas B, Forestier G, Pelletier C, Schmidt DF, Weber J, Webb GI, Idoumghar L, Muller PA, Petitjean F. InceptionTime: Finding AlexNet for time series classifica- tion, December 2020. ArXiv:1909.04939

    work page arXiv 2020

  12. [12]

    Learning from noisy labels with deep neural networks: a survey, March

    Song H, Kim M, Park D, Shin Y , Lee JG. Learning from noisy labels with deep neural networks: a survey, March

    work page

  13. [13]

    Address for correspondence: Elias Stenhede Sykehusveien 25, 1478 Nordbyhagen, Norway elias.stenhede@ahus.no

    ArXiv:2007.08199. Address for correspondence: Elias Stenhede Sykehusveien 25, 1478 Nordbyhagen, Norway elias.stenhede@ahus.no

    work page arXiv 2007