Federated Learning in Singapore Hospitals: Why Calibration Fails the Vulnerable Site

Federated Learning in Singapore Hospitals: Why Calibration Fails the Vulnerable Site

Federated learning has become the default answer when Singapore hospital clusters discuss multi-site AI deployment without centralizing patient data. Train locally, aggregate globally, preserve privacy—the pitch is compelling. But a preprint published last week on real multi-institutional brain tumor segmentation data reveals an uncomfortable truth: the standard federated calibration approach systematically harms smaller, resource-constrained sites while protecting larger ones [3]. For hospital CIOs and clinical AI teams planning federated deployments in 2026, this is not an edge case—it is a governance failure baked into the architecture.

This post is for hospital IT leaders, clinical informatics teams, and AI engineers in Singapore and Asia evaluating federated learning projects, particularly those involving multi-institutional collaborations where site size, case mix, and technical capacity vary significantly.

Key takeaways

• Pooled calibration in federated learning creates systematic harm: Standard federated conformal risk control pools calibration scores across all sites into a single threshold, causing smaller hospitals to experience 2–3× higher false negative rates than larger sites on the same model [3].
• The vulnerable site is predictable: Institutions with fewer calibration samples, different patient populations, or lower-quality imaging systematically lose coverage guarantees that larger sites retain [3].
• Human networks matter more than neural networks: Recent peer-reviewed evidence confirms that federated learning success depends more on trust, governance alignment, and institutional relationships than on algorithmic sophistication [7].
• Singapore's governance frameworks do not yet address federated fairness: The Model AI Governance Framework [1] and WHO AI ethics guidance [2] emphasize transparency and accountability but lack operational guidance for multi-site calibration equity.
• Risk-curve shrinkage offers a technical fix, but governance is the real gap: The preprint proposes site-specific risk-curve shrinkage to restore fairness, but deployment requires explicit institutional agreements on acceptable risk distribution [3].

Why federated learning calibration matters for Singapore hospitals

Federated learning is attractive for Singapore's restructured hospital clusters because it promises to train AI models across institutions without moving patient data—critical for PDPA compliance and institutional data sovereignty. The National University Health System, SingHealth, and National Healthcare Group have explored federated approaches for imaging, risk prediction, and clinical decision support.

But most federated learning tutorials and vendor pitches focus on training federation—how to aggregate model weights without centralizing data. Calibration—the step that converts raw model outputs into clinically actionable predictions with known error rates—receives far less attention. In practice, teams often pool calibration scores from all sites, compute a single threshold, and deploy it everywhere. This feels fair: everyone contributes data, everyone uses the same threshold.

The problem, quantified for the first time on real multi-institutional data in the June 2026 preprint, is that pooled calibration systematically favors larger sites [3]. On the FeTS-2022 brain tumor segmentation dataset (1,251 subjects across 20 institutions), smaller hospitals experienced false negative rates 2–3× higher than larger sites when using the pooled threshold. The model met its aggregate coverage guarantee across all sites combined, but individual vulnerable sites lost the statistical protection that conformal prediction is supposed to provide.

For a Singapore hospital deploying federated AI for stroke segmentation, sepsis prediction, or radiology triage, this is not a statistical curiosity—it is a patient safety and equity issue. If your institution is the smaller partner in a federated collaboration, your patients may receive systematically worse predictions than patients at the larger academic medical center, using the same model, under the same governance framework.

What the FeTS-2022 evidence shows

Conformal risk control (CRC) is a distribution-free method for providing statistical guarantees on prediction quality. In medical imaging, it works by calibrating a prediction-set threshold on held-out data so that the model's segmentation quality meets a target risk level (e.g., Dice score ≥ 0.7) with high probability [3].

In federated deployments, the standard approach pools calibration scores from all sites, computes a single threshold, and deploys it everywhere. The FeTS-2022 analysis shows this fails:

• Smaller sites lose coverage: Institutions contributing fewer calibration samples experienced false negative rates far above the nominal risk level, while larger sites remained well-calibrated [3].
• The effect is not random: Sites with different patient populations, imaging protocols, or scanner types systematically diverged from the pooled threshold [3].
• Aggregate metrics hide the problem: The federated model met its overall performance target across all sites combined, so standard validation would not flag the issue [3].

The preprint proposes risk-curve shrinkage—a method that adjusts each site's threshold toward the pooled estimate based on sample size and empirical variance, similar to hierarchical Bayesian shrinkage. On FeTS-2022, this restored fairness without sacrificing aggregate performance [3].

But the technical fix is only half the solution. The deeper issue is governance: who decides what "fair" means in a federated deployment? Should all sites have equal false negative rates, equal false positive rates, or equal overall accuracy? Should smaller sites accept higher error rates in exchange for access to a model they could not train alone? These are not statistical questions—they are institutional negotiations.

Why human networks matter more than neural networks

A June 2026 peer-reviewed commentary in the Journal of the American Medical Informatics Association argues that federated learning's success depends more on trust, governance alignment, and institutional relationships than on algorithmic sophistication [7]. The authors, drawing on European federated learning deployments, observe that technical failures are rare; governance failures are common.

Key failure modes include:

• Misaligned incentives: One institution wants a research publication; another wants a production clinical tool. The federated training succeeds, but the model is never deployed [7].
• Unequal contribution, unequal benefit: A smaller hospital contributes rare cases critical for model generalization but lacks the IT infrastructure to deploy the resulting model, while the larger academic partner publishes and commercializes [7].
• Calibration as an afterthought: Teams focus on training federation and treat calibration as a post-processing step, missing the fairness implications documented in [3].

For Singapore hospitals, this aligns with our experience deploying clinical AI services across restructured clusters. The hardest problems are not technical—they are governance, contracting, and institutional alignment. A federated learning project that does not explicitly address calibration fairness, benefit-sharing, and deployment equity will likely fail before the model reaches production, regardless of its AUC or Dice score.

What Singapore's governance frameworks say (and don't say)

Singapore's Model AI Governance Framework, updated by PDPC and IMDA, emphasizes transparency, explainability, and accountability [1]. It provides a structured approach to AI risk assessment and encourages organizations to document model limitations and monitor post-deployment performance.

The WHO's ethics and governance guidance for AI in health adds principles of equity, inclusiveness, and responsiveness to the needs of vulnerable populations [2]. It explicitly calls for AI systems to be designed to reduce, not exacerbate, health inequities.

But neither framework provides operational guidance for multi-site AI deployments where calibration fairness is a design choice, not an emergent property. The Model AI Governance Framework assumes a single deploying organization; the WHO guidance emphasizes principles but not implementation.

The NIST AI Risk Management Framework offers a more granular risk taxonomy, including "allocation of resources and opportunities" as a harm category [4], but it does not address federated learning specifically.

For Singapore hospital clusters planning federated AI projects in 2026, the governance gap is clear: you need an explicit institutional agreement on calibration fairness before training begins, and existing frameworks do not tell you how to structure that agreement.

A practical governance checklist for federated learning projects

Based on the evidence in [3] and [7], and our work with institutional partners in Singapore, we recommend the following governance checklist for federated learning projects involving multiple hospitals:

Before training begins

1. Define fairness explicitly: Will all sites have equal false negative rates, equal false positive rates, or equal overall accuracy? Document the choice and the rationale.
2. Quantify site heterogeneity: Measure differences in patient population, imaging protocols, and sample size across sites. If heterogeneity is high, pooled calibration will likely fail smaller sites.
3. Agree on benefit-sharing: Who owns the model? Who can deploy it? Who can publish? Who bears liability? Document this in a formal agreement before data touches the federation.
4. Plan for site-specific calibration: Budget time and data for per-site calibration or risk-curve shrinkage. Do not treat calibration as a post-processing afterthought.

During training and validation

1. Monitor per-site performance: Track calibration quality, false negative rates, and false positive rates per site, not just in aggregate. If smaller sites diverge, stop and re-calibrate.
2. Test on held-out sites: If possible, reserve one or two sites entirely for testing. This reveals generalization failures that cross-validation within the federation will miss.

Before deployment

1. Require per-site sign-off: Each deploying institution should review its own site-specific performance and explicitly approve deployment. Do not deploy a model to a site that did not consent based on its own data.
2. Document calibration methodology: Record whether you used pooled calibration, per-site calibration, or risk-curve shrinkage. This is a governance artifact, not just a technical detail.

Post-deployment

1. Monitor drift per site: Federated models can drift differently at different sites due to local changes in patient population, imaging equipment, or clinical workflows. Per-site monitoring is essential (see our post on clinical AI drift monitoring).
2. Re-calibrate periodically: Calibration is not a one-time step. Plan for periodic re-calibration using fresh data from each site.

Why this matters in Singapore and Asia

Singapore's restructured hospital clusters are natural candidates for federated learning: they share governance structures, clinical protocols, and IT standards, but patient data remains institutionally siloed. Federated learning offers a path to multi-site AI without centralized data lakes.

But Singapore's hospitals vary significantly in size, case mix, and technical capacity. A federated model trained across NUHS, SingHealth, and NHG will likely favor the largest sites unless calibration fairness is explicitly designed in. Smaller community hospitals, which serve older and more socioeconomically vulnerable populations, risk becoming the "vulnerable site" in [3].

Across Asia, federated learning is increasingly positioned as a solution for cross-border AI collaboration—training models across Singapore, Malaysia, Thailand, and Indonesia without moving data across borders. The calibration fairness problem is even more acute in this context: differences in healthcare infrastructure, imaging quality, and patient populations are larger, and governance alignment is harder.

The June 2026 evidence from FeTS-2022 [3] and the JAMIA commentary [7] make clear that federated learning is not a purely technical problem. It is a governance problem that requires explicit institutional agreements on fairness, benefit-sharing, and calibration methodology.

What to do next

If you are planning a federated learning project in a Singapore hospital or health system:

• Read the FeTS-2022 preprint [3] and discuss the calibration fairness problem with your AI team and clinical partners before training begins.
• Use the governance checklist above to structure your institutional agreements. Do not start training until you have explicit consensus on calibration fairness and benefit-sharing.
• Budget for per-site validation and calibration: Treat this as a governance cost, not a technical overhead. It is cheaper than deploying a model that harms vulnerable sites.
• Engage with Singapore's Model AI Governance Framework [1] and adapt it for multi-site deployments. Document your calibration methodology as part of your AI risk assessment.
• Consider starting a conversation with InsytAI if you need help structuring governance for a federated learning project or evaluating calibration fairness in an existing deployment.

FAQ

What is conformal risk control, and why does it matter for federated learning?

Conformal risk control (CRC) is a statistical method that provides distribution-free guarantees on prediction quality by calibrating a threshold on held-out data [3]. In federated learning, the standard approach pools calibration data from all sites, but this systematically harms smaller sites. CRC matters because it makes the fairness problem quantifiable and addressable.

Can we just use per-site calibration instead of pooled calibration?

Yes, but per-site calibration requires enough data at each site to be statistically reliable. For smaller hospitals, this may not be feasible. Risk-curve shrinkage [3] offers a middle ground: it adjusts each site's threshold toward the pooled estimate based on sample size, balancing fairness and statistical efficiency.

Does Singapore's Model AI Governance Framework address federated learning?

The Model AI Governance Framework [1] provides principles for transparency, explainability, and accountability, but it does not specifically address multi-site AI deployments or calibration fairness. Hospital clusters planning federated projects need to extend the framework with explicit governance agreements on per-site performance and benefit-sharing.

Is federated learning still worth doing after this evidence?

Yes, but with eyes open. Federated learning remains the best approach for multi-site AI when data cannot be centralized. The FeTS-2022 evidence [3] does not argue against federated learning—it argues for better calibration methodology and explicit governance of fairness tradeoffs. The technical fix (risk-curve shrinkage) is straightforward; the governance fix (institutional agreements on fairness) is harder but essential.

Sources

[1] Singapore Model AI Governance Framework — PDPC Singapore. https://www.pdpc.gov.sg/help-and-resources/2020/01/model-ai-governance-framework

[2] WHO ethics and governance of artificial intelligence for health — WHO. https://www.who.int/publications/i/item/9789240029200

[3] When Calibration Fails the Vulnerable Hospital: Federated Conformal Risk Control via Risk-Curve Shrinkage — arXiv cs.LG+clinical 2026-06-18. https://arxiv.org/abs/2606.20115v1

[4] NIST AI Risk Management Framework — NIST. https://www.nist.gov/itl/ai-risk-management-framework

[7] Federated learning's uncomfortable truth: why human networks matter more than neural networks — Journal of the American Medical Informatics Association : JAMIA 2026 Jun 1. https://pubmed.ncbi.nlm.nih.gov/41984621/