Medical Imaging Foundation Models: Predictability Bias in Singapore Hospitals

Medical Imaging Foundation Models: Predictability Bias in Singapore Hospitals

A new wave of medical imaging foundation models promises to reduce annotation burden and improve generalization across tasks. But recent research reveals a fundamental architectural constraint that Singapore hospitals need to understand before deploying these systems: predictability bias in joint-embedding predictive architectures (JEPAs). If you're evaluating imaging AI for radiology, cardiology, or pathology workflows, this constraint affects what your model can and cannot learn—and it's not captured by standard validation metrics.

Key takeaways

• Joint-embedding predictive architectures (JEPAs) learn by predicting masked regions from visible context, but this creates predictability bias: models over-represent recurrent anatomy and under-represent sporadic pathology [1]
• In CT imaging, recurrent anatomical structures are predictable from context, while sporadic lesions and pathology are systematically underweighted during pre-training [1]
• New open vision-language models for medical imaging are emerging, but deployment requires testing for task-specific bias, not just aggregate accuracy [2]
• Measurement-grounded representation learning (like EchoTab for echocardiography) offers an alternative architecture that explicitly models clinical measurements before diagnosis [3]
• Singapore hospitals should validate imaging foundation models against task-specific pathology detection, not just reconstruction loss or generic segmentation benchmarks

What is predictability bias in medical imaging foundation models?

Joint-embedding predictive architectures have become popular for self-supervised learning in medical imaging. The idea is elegant: mask portions of an image, then train the model to predict the representation of masked regions from visible context. This approach has worked well in natural images, where context often determines content.

But medical imaging is different. A recent preprint analyzing JEPAs in computed tomography reveals a critical constraint [1]: point prediction summarizes the conditional target distribution and cannot reproduce variation that the context does not resolve. In plain English: if a lesion could appear anywhere in the liver, and the visible context doesn't tell you where, the model learns to predict the average—which is usually "no lesion."

The result is predictability bias. Recurrent anatomy (ribs, vessels, organ boundaries) is highly predictable from context, so the model learns rich representations of normal structure. Sporadic pathology (tumors, hemorrhages, fractures) is not predictable from context, so the model systematically underweights these features during pre-training.

This isn't a bug—it's a fundamental property of the architecture. And it has direct implications for downstream task performance, especially for rare or spatially variable pathology.

Why standard validation metrics miss this problem

Most foundation model papers report aggregate metrics: reconstruction loss, segmentation Dice scores, classification AUC across a held-out test set. These metrics can look excellent even when predictability bias is present, because:

1. Reconstruction loss is dominated by high-frequency anatomical structure, which JEPAs learn well
2. Segmentation benchmarks often focus on large, consistent structures (organs, vessels) rather than sporadic lesions
3. Classification AUC aggregates across all cases, masking poor performance on rare pathology subtypes

A model can achieve 95% AUC on a lung nodule detection task while systematically missing ground-glass opacities or subsolid nodules—because these features are less predictable from context and therefore underweighted during pre-training.

For Singapore hospitals evaluating imaging foundation models, this means you need task-specific validation on the pathology you care about, not just the vendor's aggregate benchmark results. If you're deploying for stroke detection, validate on hemorrhage subtypes. If you're deploying for oncology, validate on early-stage lesions and unusual presentations.

Alternative architectures: measurement-grounded representation learning

Not all foundation model architectures have this constraint. A new preprint on echocardiography introduces EchoTab, which learns representations by reconstructing clinical measurements rather than predicting masked image regions [3].

The architecture is straightforward: the model learns to map multi-view echocardiographic images to a tabular representation of quantitative measurements (ejection fraction, chamber dimensions, valve parameters), which then support specific diagnoses. This mirrors clinical workflow—cardiologists interpret echo studies through measurements, not raw pixel prediction.

Because the model is trained to predict explicit measurements, it must learn features that correlate with those measurements, even if they're not predictable from spatial context alone. A dilated left ventricle has a measurable dimension; the model learns to detect it because that measurement is part of the training objective, not because it's spatially predictable.

This approach is particularly relevant for Singapore hospitals deploying imaging AI in cardiology, where quantitative measurements are already part of structured reporting workflows. If your PACS and reporting system already capture structured measurements, measurement-grounded architectures align with existing clinical process and governance requirements.

What about vision-language models for medical imaging?

A new open vision-language model for medical imaging was published in Nature Medicine this month [2]. These models learn joint representations of images and text (radiology reports, clinical notes), enabling zero-shot and few-shot task adaptation without task-specific fine-tuning.

Vision-language models avoid some of the predictability bias issues of JEPAs, because the text supervision provides explicit signal about pathology presence and characteristics. If the radiology report describes a "3 cm hypodense lesion in segment 7 of the liver," the model learns to associate that text with the corresponding image features, even if the lesion location wasn't predictable from anatomical context.

But vision-language models introduce different deployment challenges for Singapore hospitals:

• Text quality matters: if your radiology reports use inconsistent terminology, abbreviations, or incomplete descriptions, the model's text encoder won't generalize well
• Language coverage: most vision-language models are trained on English text; if your institution uses Mandarin, Malay, or Tamil in clinical notes, you'll need multilingual models or translation pipelines
• Prompt engineering: zero-shot performance depends heavily on how you phrase the query; this requires clinical validation, not just engineering experimentation

For Singapore hospitals, vision-language models are most promising for cross-institutional generalization (where training data comes from international sources but deployment is local) and rare disease detection (where you have few local examples but can leverage text descriptions). But you still need local validation on your patient population and reporting style.

Why this matters in Singapore

Singapore's public healthcare clusters are actively deploying medical imaging AI across radiology, cardiology, and pathology workflows. The shift from task-specific models to foundation models is attractive: lower annotation burden, better generalization, faster adaptation to new tasks.

But foundation models are not plug-and-play. Predictability bias, measurement grounding, and vision-language alignment are architectural choices with direct clinical implications. A model that performs well on international benchmarks may underperform on the specific pathology subtypes, patient demographics, or imaging protocols used in Singapore hospitals.

This is especially relevant for HSA AI-SaMD pathways. If you're deploying a foundation model under the HSA AI-SaMD exemption pathway, you need to demonstrate that the model performs adequately on the specific clinical use case in your institution, not just on aggregate benchmarks. Predictability bias is a plausible mechanism for performance degradation that won't show up in vendor validation reports.

For clinical AI teams building platforms for clinical AI services, foundation models offer a useful starting point for feature extraction—but you should validate representations on your downstream task before committing to a specific architecture. If your task involves detecting rare or spatially variable pathology, measurement-grounded or vision-language architectures may generalize better than JEPAs.

What to do next

• Audit your validation strategy: if you're evaluating an imaging foundation model, test performance on rare pathology subtypes and unusual presentations, not just aggregate metrics
• Request task-specific validation data: ask vendors for per-class performance breakdowns, especially for low-prevalence conditions and edge cases
• Consider measurement-grounded architectures if your clinical workflow already captures structured quantitative measurements (echo, cardiac MRI, bone density)
• Pilot vision-language models for cross-institutional generalization, but validate prompt engineering and text quality on your local radiology reports
• Document architectural constraints in your model risk assessment: predictability bias is a known limitation of JEPAs; if you're deploying one, document how you've validated performance on non-recurrent pathology

If you're building or deploying medical imaging AI in Singapore and need help designing task-specific validation protocols or evaluating foundation model architectures, start a project with us.

FAQ

What is a joint-embedding predictive architecture (JEPA)?

A JEPA is a self-supervised learning method that trains a model to predict the representation of masked image regions from visible context. It's popular for foundation models because it doesn't require labels. But in medical imaging, this creates predictability bias: the model learns to represent recurrent anatomy well and sporadic pathology poorly [1].

Should Singapore hospitals avoid JEPAs for medical imaging?

Not necessarily. JEPAs work well for tasks where the target pathology is spatially consistent or predictable from anatomical context (e.g., organ segmentation, vessel detection). But for rare or spatially variable pathology (e.g., early-stage tumors, hemorrhage subtypes), you should validate carefully and consider alternative architectures like measurement-grounded or vision-language models.

How do I validate a foundation model for task-specific pathology detection?

Request per-class performance metrics from the vendor, especially for low-prevalence conditions. Run your own validation on a local test set that includes edge cases, unusual presentations, and rare subtypes. Compare performance to a task-specific model trained on your data. If the foundation model underperforms on specific pathology classes, document this in your model risk assessment and consider fine-tuning or alternative architectures.

What are measurement-grounded representations and when should I use them?

Measurement-grounded models (like EchoTab [3]) learn to predict clinical measurements (ejection fraction, chamber dimensions) rather than reconstruct masked image regions. This approach works well when your clinical workflow already captures structured measurements and when pathology is defined by quantitative thresholds rather than spatial patterns. It's particularly relevant for cardiology, musculoskeletal imaging, and bone density assessment.

Sources

[1] The Predictability Bias of Joint-Embedding Predictive Learning in Medical Imaging. arXiv cs.LG+clinical. 2026 Oct 9. https://arxiv.org/abs/2610.13237v1

[2] Sellergren A, Kazemzadeh S, Mahvar F. An open vision-language model for diverse medical applications. Nature Medicine. 2026 Oct 6. https://pubmed.ncbi.nlm.nih.gov/42839104/

[3] EchoTab: Learning Measurement-Grounded Representations for Echocardiographic Diagnosis through Tabular Reconstruction. arXiv cs.LG+clinical. 2026 Oct 9. https://arxiv.org/abs/2610.13599v1

[4] Adapting Segment Anything Model 3 for concept-driven lesion segmentation in medical images: an experimental study. Visual computing for industry, biomedicine, and art. 2026 Oct 1. https://pubmed.ncbi.nlm.nih.gov/42862513/

[5] Deep learning-based segmentation, grading and genotype classification of glioma: a literature analysis. Quantitative imaging in medicine and surgery. 2026 Oct 1. https://pubmed.ncbi.nlm.nih.gov/42861085/