REVIEW 2 major objections 2 minor
Biases introduced during omics data collection get locked into biomedical foundation models and produce downstream healthcare inequities that later rules cannot fix.
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 →
T0 review · grok-4.3
2026-05-10 11:47 UTC pith:MKH4PNAP
load-bearing objection This perspective gives concrete counts on ancestry underreporting in omics papers and datasets but assumes rather than demonstrates that the bias creates irreversible downstream healthcare disparities. the 2 major comments →
Perspective on Bias in Biomedical AI: Preventing Downstream Healthcare Disparities
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
As biomedical foundation models become central to discovery through repeated reuse of models pretrained on large omics collections, the documented underreporting of ancestry and strong European dominance in those collections will be perpetuated and amplified, producing performance gaps and health inequities for non-European populations that regulatory interventions at later stages cannot fully reverse.
What carries the argument
The pretraining-and-reuse paradigm for foundation models, which transfers population skews present in source omics datasets into every downstream application.
Load-bearing premise
That the observed dominance of European-ancestry samples in omics datasets will produce measurable differences in model accuracy or clinical outcomes for other ancestry groups.
What would settle it
A controlled experiment that trains two otherwise identical foundation models, one on current European-heavy omics data and one on a version balanced across ancestries, then measures no difference in downstream task performance or fairness metrics on held-out non-European cohorts.
If this is right
- Regulatory interventions applied only at clinical deployment will leave early-stage data biases intact.
- Community adoption of Provenance, Openness, and Evaluation Transparency practices would reduce the risk of irreversible inequities.
- Biomedical AI tools will serve underserved populations more effectively once demographic composition of training data is routinely disclosed and evaluated.
- Repeated reuse of the same biased base models across tasks will compound rather than dilute the initial population skew.
Where Pith is reading between the lines
- Future work could test whether adding even modest amounts of non-European omics data at the pretraining stage measurably improves equity metrics without harming overall accuracy.
- The same logic may apply to other data modalities such as imaging or electronic health records that feed into shared foundation models.
- Funding agencies could require ancestry reporting as a condition for dataset deposition to change collection incentives upstream.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript is a perspective arguing that biases arise early in biomedical research during omics data collection and prioritization. An automated analysis of 4719 PubMed-indexed omics publications (2015-2024) finds low rates of ancestry/ethnicity reporting (with modest improvement over time), while inspection of CellxGene and GEO datasets shows strong European-ancestry dominance. The authors contend that, under the foundation-model paradigm of large-scale pretraining followed by repeated downstream reuse, these early biases will be perpetuated or amplified, producing cascading healthcare disparities that later regulatory interventions cannot fully reverse. They advocate three community principles—Provenance, Openness, and Evaluation Transparency—to improve equity and robustness.
Significance. If the causal pathway from dataset demographics to irreversible downstream disparities is substantiated, the perspective would usefully direct attention to upstream data practices in biomedical AI. The concrete counts from the 4719-publication corpus and the two large public repositories supply a tangible empirical anchor for the bias observation, which is a clear strength. The proposed principles offer a practical, non-regulatory framing that could influence data-sharing norms and model documentation standards.
major comments (2)
- [Abstract] Abstract and the paragraph introducing the foundation-model paradigm: the central claim that pretraining on ancestry-biased omics data will produce 'cascading inequities that regulatory interventions cannot fully reverse' is asserted without direct empirical support or simulation inside the manuscript. The 4719-publication counts and CellxGene/GEO inspections establish the existence of reporting gaps and population imbalance, but no biomedical-specific evidence, ablation, or outcome-linked analysis demonstrates that these translate into ancestry-linked performance gaps in foundation models or into health inequities immune to later mitigation.
- [Foundation-model risk discussion] Section discussing risks to downstream tasks: the mechanism by which European dominance in pretraining corpora is expected to propagate into measurable disparities for non-European populations in clinical AI applications is described at a high level but not instantiated with any concrete example, performance metric, or reference to a controlled study within the paper, leaving the load-bearing causal step untested.
minor comments (2)
- [Automated analysis of publications] The automated-analysis subsection would benefit from explicit reporting of the PubMed query string, the exact criteria or classifier used to flag ancestry mentions, and any validation steps (e.g., manual review of a sample), which are necessary for reproducibility of the 4719-paper statistics.
- [Conclusion] The manuscript would be strengthened by a brief discussion of how the three proposed principles (Provenance, Openness, Evaluation Transparency) could be operationalized in existing data repositories or model cards, moving from high-level recommendation to actionable guidance.
Simulated Author's Rebuttal
We thank the referee for the constructive comments. We address each major point below and have revised the manuscript to qualify our claims more carefully, add supporting literature citations, and expand the discussion of mechanisms while preserving the perspective's focus on upstream data practices.
read point-by-point responses
-
Referee: [Abstract] Abstract and the paragraph introducing the foundation-model paradigm: the central claim that pretraining on ancestry-biased omics data will produce 'cascading inequities that regulatory interventions cannot fully reverse' is asserted without direct empirical support or simulation inside the manuscript. The 4719-publication counts and CellxGene/GEO inspections establish the existence of reporting gaps and population imbalance, but no biomedical-specific evidence, ablation, or outcome-linked analysis demonstrates that these translate into ancestry-linked performance gaps in foundation models or into health inequities immune to later mitigation.
Authors: We agree that the manuscript, as a perspective, does not include original empirical simulations, ablations, or outcome-linked analyses demonstrating the full causal translation from biased pretraining data to irreversible downstream disparities. Our contribution centers on documenting the upstream imbalances via the PubMed corpus analysis and repository inspections, then linking these to the foundation-model reuse paradigm. In revision, we have softened the abstract and introduction to describe a 'risk of perpetuating or amplifying biases, potentially leading to cascading inequities that may prove difficult to fully reverse through later interventions alone.' We have also added citations to studies documenting ancestry-linked performance gaps in genomic and single-cell AI models to provide indirect support for the mechanism. revision: yes
-
Referee: [Foundation-model risk discussion] Section discussing risks to downstream tasks: the mechanism by which European dominance in pretraining corpora is expected to propagate into measurable disparities for non-European populations in clinical AI applications is described at a high level but not instantiated with any concrete example, performance metric, or reference to a controlled study within the paper, leaving the load-bearing causal step untested.
Authors: We acknowledge that the original discussion of propagation remained conceptual. The revised manuscript expands this section with concrete examples and references drawn from the literature, including documented reductions in accuracy for polygenic risk scores and variant interpretation models when applied to non-European ancestry groups after European-dominant pretraining, as well as ancestry biases observed in cell-type annotation from single-cell omics foundation models. These additions instantiate the mechanism with specific performance considerations while clarifying that the degree of irreversibility depends on the feasibility of downstream mitigation. revision: yes
Circularity Check
No circularity: claims rest on external dataset analysis and logical inference, not self-referential derivations.
full rationale
The paper performs an automated count of ancestry reporting in 4719 PubMed omics papers (2015-2024) and inspects demographic composition in CellxGene and GEO. It then reasons that foundation-model pretraining on such data may perpetuate biases into downstream tasks. This is observational reporting plus perspective, with no equations, fitted parameters, self-defined terms, or load-bearing self-citations that reduce the central claim to its own inputs by construction. The causal extrapolation to irreversible inequities is an interpretive step, not a mathematical reduction. No steps match the enumerated circularity patterns.
Axiom & Free-Parameter Ledger
axioms (2)
- domain assumption Biases present at data collection will be perpetuated or amplified when models are pretrained on large omics datasets and reused for downstream tasks.
- ad hoc to paper Regulatory interventions cannot fully reverse early-stage biases once embedded in foundation models.
read the original abstract
Healthcare disparities persist across socioeconomic boundaries, often attributed to unequal access to screening, diagnostics, and therapeutics. However, this perspective highlights that critical biases can emerge much earlier, during data collection and research prioritization, long before clinical implementation, particularly in studies focused on molecular and omics data. A vast number of studies focus on collecting omics data, but the demographic information associated with these datasets is often not reported, and when it is reported, it reveals substantial biases. An automated analysis of 4514 PubMed-indexed omics publications from 2015 to 2024, examining reporting across multiple demographic dimensions, reveals limited reporting overall; for example, only 2.7% of studies report ancestry or ethnicity information and geographic origin reporting is limited to 2.5%. Analysis of large-scale datasets commonly used for model training, such as CellxGene and GEO, reveals substantial population bias where European-ancestry data dominates. As biomedical foundation models become central to biomedical discovery with a paradigm in which base models are pretrained on large datasets and reusing them repeatedly for many different downstream tasks, they risk perpetuating or amplifying these early-stage biases, leading to cascading inequities that regulatory interventions cannot fully reverse. We propose a community-wide focus on three foundational principles: Provenance, Openness, and Reliability through Evaluation Transparency. Together, these principles can help make biases and limitations more visible to model developers and users, supporting more informed model development, evaluation, and deployment decisions in biomedical AI.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.