a Singapore tertiary system deploys diagnostic agents.

The system and the mandate

Singapore's tertiary healthcare structure is unusual by regional standards: two integrated clusters, SingHealth and NUHS, serve a resident population of 5.9 million across a combined estate of eleven acute hospitals, seven national specialty centres, and a polyclinic network that interfaces directly with the Ministry of Health's Healthier SG primary care reform programme, launched in 2023. Healthier SG is not an incremental policy adjustment. It is a structural redesign of how the Singapore health system routes patients — enrolling residents with a single primary care provider and mandating that tertiary facilities receive referrals through a defined gatekeeping interface. That interface, a longitudinal health record visible to both primary care providers and tertiary teams, became the architectural foundation on which the diagnostic agent deployment was built.

Dr. Tan Wei Liang, Chief Medical Information Officer of the composite cluster since April 2022, set the terms of the programme in an internal mandate that circulated to department heads in September 2022. The mandate was specific in two ways that distinguished it from the wave of clinical AI press releases arriving from American health systems at the time. First, it committed the cluster to a vendor-neutral evaluation architecture — no single vendor would supply the diagnostic layer; the cluster would evaluate competing capabilities against a shared framework and select on performance, not partnership. Second, it required that any diagnostic agent deployed within the cluster's acute facilities operate effectively across all four of Singapore's official languages. This was not aspirational language. It was a technical specification, and it eliminated most of the vendors who responded to the initial market consultation in late 2022.

The multi-language requirement reflects a clinical reality that Singapore's health system navigates daily. Elderly Mandarin-speaking patients at tertiary outpatient clinics routinely describe symptoms in dialect-inflected Mandarin that does not map cleanly to standard simplified Chinese clinical terminology. Tamil-speaking patients — a significant portion of the community served by the National University Hospital and Ng Teng Fong General Hospital — present linguistic patterns that differ materially from training data drawn from Indian clinical populations. Malay-language clinical input arrives in a register shaped by code-switching with English that no model trained on monolingual Malay corpora handles reliably. Tan's team documented 47 distinct failure modes across these language populations during the evaluation phase. The vendor field narrowed from 21 expressions of interest to three capable of addressing more than 30 of them.

Vendor selection and the PRISM framework

The evaluation framework the cluster developed — designated the Polylingual and Risk-stratified Inference Standards Matrix, PRISM — is the most operationally specific clinical AI evaluation architecture published in Southeast Asia. Developed by Tan's office in collaboration with MOH Holdings' technology procurement unit and the Agency for Science, Technology and Research's Centre for Frontier AI Research, PRISM runs candidate diagnostic agent capabilities through five evaluation gates before any system receives conditional deployment authority. Gate one is benchmark performance: agents evaluated against a gold-standard case set of 3,100 de-identified patient records drawn from the cluster's own clinical data repository, curated to reflect the full demographic and linguistic distribution of the cluster's inpatient population. No vendor may substitute their own evaluation dataset. Gate two is polylingual consistency: the same clinical case, presented identically in all four languages, must produce outputs that a bilingual clinical review panel judges diagnostically equivalent at a threshold of 91 per cent. Gate three is adverse-event sensitivity: agents are tested against a set of 180 cases in which a diagnostic error would carry high clinical consequence — sepsis, acute coronary syndrome, stroke, paediatric fever with atypical presentation — using a subtraction battery that removes individual data elements to test for catastrophic versus graceful output degradation. Gate four is Healthier SG interface fidelity: agents must retrieve and interpret the longitudinal health record correctly in 98 per cent of test cases, including records where data entry discipline across contributing primary care providers is inconsistent. Gate five is MOH regulatory review: the completed evaluation dossier is submitted to the Ministry of Health's Health Products Regulation Group for assessment under Singapore's Software as a Medical Device guidance, updated in November 2023 to explicitly cover multi-step diagnostic agent capabilities.

Three vendors cleared all five gates. Holmium Health, a Singapore-incorporated clinical AI company with research roots in the National University of Singapore's Yong Soo Loo School of Medicine, supplied the differential diagnosis agent for acute and emergency settings. Arceden Systems, a Taiwanese vendor with a Southeast Asia commercial presence since 2021, supplied the radiology triage and report structuring layer. The third vendor — an unnamed Singapore government-linked technology entity operating through MOH Holdings' digital health subsidiary — supplied the Healthier SG interface integration layer that connects the diagnostic agents to the primary care longitudinal record. Tan's team has declined to name the third vendor in any public communication; the February 2024 procurement notice refers to it only as "Supplier C."

Holmium Health's selection was the most contested element of the vendor award. Three international vendors — including at least one with an existing FDA clearance portfolio — challenged the PRISM gate two results in a procurement review filed with the Government Procurement (Amendment) Act complaints mechanism in November 2023. The review panel, chaired by a retired Subordinate Courts judge, found that the gate two polylingual consistency threshold had been correctly applied and that the challengers' submissions had not met it. The review outcome was published in January 2024. It established, for the first time in Singapore procurement law, a precedent that language-specific clinical performance is a legitimate technical evaluation criterion rather than a preference that must yield to price-point considerations.

MOH regulatory dialogue: cooperative, structured, and ahead of the curve

Singapore's Ministry of Health began its formal engagement with the Health Products Regulation Group in Q1 2023, eleven months before the procurement award. The cluster's approach mirrors the posture Mayo Clinic adopted with the FDA — voluntary, proactive, and structured around the assumption that early regulatory alignment is less expensive than retroactive compliance — but adapts it to Singapore's distinctly different regulatory environment. Singapore's HPRG operates under the Health Products Act, which was amended in 2021 to bring software as a medical device under its licensing framework. The November 2023 update to the HPRG's SaMD guidance added a specific classification track for "conversational and multi-step clinical reasoning software," covering capabilities that go beyond discrete algorithmic outputs to engage in extended diagnostic workflows. The cluster's PRISM framework was developed in direct consultation with the HPRG's Medical Technology Standards team, and the gate five dossier structure was designed to map directly onto the HPRG's SaMD technical documentation requirements.

Dr. Noor Aisha Binte Rahmat, Deputy Director of the HPRG's Digital Health Office, confirmed in an MOH briefing note published alongside the November 2023 guidance update that the cluster's evaluation methodology had been reviewed as a reference case during the drafting of the new classification track. The briefing note stopped short of endorsing PRISM as a regulatory standard — the HPRG does not pre-endorse private evaluation frameworks — but its language made the alignment clear. Tan's team read the briefing note as a signal that the gate five dossiers, when submitted in January 2024, would not face the interpretive gaps that had extended the HPRG review timeline for earlier-generation clinical AI submissions by two to three quarters. The dossiers for Holmium Health and Arceden Systems received conditional approval notifications in 21 and 28 days respectively — well inside the HPRG's published 90-day indicative timeline.

The regulatory dialogue also addressed the question of liability allocation — a pressure point in every clinical AI governance conversation — with more specificity than comparable frameworks in the region have achieved. The HPRG's November 2023 guidance established that clinical accountability for any output produced by a conditionally approved diagnostic agent remains with the attending clinician, subject to the condition that the deploying institution can demonstrate the agent was operating within its validated scope at the time the output was generated. That demonstration requires the audit trail architecture that Tan's team built into the PRISM framework from the first design iteration: a structured interaction log capturing the agent's input context, output confidence tier, and any refusal triggers activated during the encounter, retained in the cluster's electronic medical record system for a minimum of seven years and accessible for HPRG audit within 72 hours of request.

The Healthier SG interface: agents at the primary-tertiary junction

Healthier SG is the operational context that makes Singapore's diagnostic agent deployment strategically distinct from comparable programmes at Mayo, the NHS, or the German academic medical centres beginning to publish deployment notes. The reform programme, which enrolled its first 300,000 residents in October 2023 and targets full resident enrolment by 2030, creates a longitudinal health record that follows the patient from the enrolled primary care provider through every tertiary interaction. That record is the input environment in which the cluster's diagnostic agents operate, and its quality — the completeness and consistency of data entered across Singapore's mixed public-private primary care landscape — determines the agents' performance in ways that no internal hospital EMR deployment faces.

The gate four evaluation in PRISM exists because of this challenge. Primary care providers contributing to the Healthier SG record include 1,300 participating private general practitioners operating across EMR systems that range from purpose-built Healthier SG-certified platforms to older legacy systems with minimal structured data fields. The cluster's evaluation found that longitudinal records from practices using non-certified legacy systems had structured data completeness rates 34 percentage points lower than records from certified-platform practices. Holmium Health's differential diagnosis agent, when tested against low-completeness records, produced output confidence tiers that dropped below the PRISM tier two advisory threshold — meaning the agent would present a differential with an explicit caveat that structured data quality was insufficient for a tier one recommendation — in 23 per cent of legacy-practice cases. That caveat is not a failure. It is the designed escalation behaviour. The cluster's clinical governance team reviewed 400 of these caveat-flagged outputs during the evaluation phase and found that in 89 per cent of cases the attending clinician's manual review of the same incomplete record produced equivalent diagnostic uncertainty. The agent was calibrated to the data, not performing below it.

The interface also carries the multi-language challenge in a different form. Primary care clinical notes contributed to the Healthier SG record arrive in the language the GP used to document the encounter — which in Singapore's primary care landscape means a distribution across English, Mandarin, and Malay that reflects the GP's own linguistic preferences, not the patient's. A Tamil-speaking elderly patient enrolled with a Mandarin-preferring GP may have their presenting complaint documented in English, their medication history in a mix of English trade names and Mandarin generic equivalents, and their family history omitted because the consultation occurred through a family member acting as interpreter whose linguistic contribution was not captured in the record. Holmium Health's handling of this scenario — which the PRISM evaluation documented as occurring in 11 per cent of test cases involving Tamil-speaking patients — was the differentiating factor in its final scoring against its closest competitor. The agent flagged language provenance uncertainty in the structured output and presented a reduced confidence tier with a specific note identifying the data gap. No other evaluated vendor produced a structured flag for language provenance inconsistency rather than a degraded output with no explanatory attribution.

What to watch

The cluster moves to phased live deployment beginning in Q2 2024, starting with emergency department triage support at Singapore General Hospital and National University Hospital before extending to acute medical admissions and outpatient specialty settings by year-end. Five variables determine whether this programme becomes the reference case for diagnostic agent governance in Southeast Asia or remains a Singapore-specific implementation.

• Whether the HPRG's November 2023 SaMD guidance update — the first in the region to explicitly classify multi-step diagnostic agents — is adopted or referenced by peer regulators in Malaysia, Thailand, and Indonesia, who are watching the Singapore framework closely; if two or more ASEAN regulators align their SaMD classification logic with Singapore's, PRISM becomes a de facto regional standard and Holmium Health's gate two performance record becomes a commercial moat in every market where the standard applies.
• Whether the polylingual consistency threshold of 91 per cent holds in live deployment across the cluster's full patient volume, particularly for Tamil-language inputs at National University Hospital and Ng Teng Fong General Hospital, where the Tamil-speaking elderly inpatient population represents the highest-risk test of a language performance threshold that was set on evaluation data and has not yet been validated at production scale.
• Whether the Healthier SG longitudinal record's structured data completeness improves at a rate that closes the 34-percentage-point gap between certified and legacy primary care systems within the next 18 months; MOH has set a certified-platform adoption target for participating GPs but has not attached financial penalties to non-compliance, and the cluster's gate four results indicate that legacy-practice data quality is the binding constraint on agent performance at the primary-tertiary interface.
• Whether "Supplier C" — the government-linked entity supplying the Healthier SG integration layer — is identified publicly and whether its technology is offered to other ASEAN public health systems through a bilateral or multilateral arrangement; Singapore's track record of offering public-sector technology through government-to-government channels makes this scenario plausible, and the integration layer is, in technical terms, the most transferable component of the deployment architecture.
• Whether the procurement review precedent established in January 2024 — that language-specific clinical performance is a legitimate technical evaluation criterion — survives a potential appeal to the Singapore International Commercial Court and, if it does, whether it influences how MOH Holdings structures future clinical AI procurement across the broader public health cluster, including the polyclinic network that will carry the highest patient volume once Healthier SG enrolment reaches full scale.

Frequently asked

What is PRISM, and is it a published standard?

PRISM — the Polylingual and Risk-stratified Inference Standards Matrix — is the evaluation framework developed by the cluster's Office of the CMIO in collaboration with MOH Holdings and A*STAR's Centre for Frontier AI Research. It is not a published standard. The framework's gate structure and threshold parameters are described in the February 2024 procurement award notice, but the full technical specification is a non-public working document. The HPRG's November 2023 SaMD guidance references the framework's methodology without naming it, which is the closest to official endorsement it has received. Several peer health systems in the region have requested the framework documentation through bilateral Ministry of Health channels; none have received the full specification as of the date of this briefing.

Who pays for the deployment, and how is the funding structured?

The deployment is funded through a combination of MOH Holdings' technology capital budget and a Singapore Ministry of Health Innovation Fund allocation announced in the Ministry's Q3 2023 budget supplement. MOH Holdings is the holding company for Singapore's public healthcare clusters, and its technology procurement function negotiated the vendor contracts on behalf of the composite cluster. The specific contract values are not publicly disclosed under the terms of the Government Procurement Act's commercial sensitivity provisions. The Healthier SG integration component — supplied by Supplier C — is funded separately through MOH's Healthier SG programme budget rather than through the cluster's technology capital allocation, which is why it does not appear in the three-vendor procurement award notice; it was commissioned under a direct government-to-subsidiary arrangement that does not require public tender under Singapore procurement rules for transactions between government-linked entities.

Why did international vendors with FDA clearance fail the evaluation?

The primary failure point was gate two: polylingual consistency at the 91 per cent threshold. FDA-cleared diagnostic agents evaluated during the PRISM process had been trained predominantly on English-language clinical data with secondary Mandarin capability developed for the Chinese-American patient population — a linguistic profile that does not transfer reliably to Singapore's Mandarin register, and which has no coverage for Malay or Tamil at clinical-grade performance levels. Gate four also eliminated several international candidates: their architectures did not support the Healthier SG longitudinal record's data schema, and the integration work required to achieve 98 per cent fidelity on the gate four test set was outside the timeline the procurement process allowed. The vendors that filed the procurement review argued that the gate two threshold was set at a level achievable only by vendors with Singapore-specific training data — an argument the review panel found was an accurate description of the threshold's effect but not evidence that the threshold was set for anti-competitive purposes.

How does the deployment handle clinical liability when an agent output is acted upon by an attending clinician?

Under the HPRG's November 2023 SaMD guidance and the cluster's internal governance framework, clinical accountability for any decision made in response to an agent output remains with the attending clinician. The agent's conditional approval is contingent on it operating within its validated scope — defined by the PRISM gate parameters — and the cluster's audit trail captures whether the agent was operating within scope at the time of each interaction. If a clinical safety incident is investigated and the audit trail shows the agent was operating outside validated scope, liability exposure shifts toward the cluster as the deploying institution and, potentially, the vendor. If the agent was operating within validated scope, the attending clinician's decision to act on the output is treated as a clinical decision under the existing professional liability framework. The cluster's legal counsel reviewed this allocation with MOH's legal division and the Singapore Medical Council before the deployment commenced. No new legislation was required; the existing framework was interpreted to accommodate it.

What does the deployment mean for clinical staff in the cluster's emergency departments?

In the emergency department triage setting — the first live deployment context beginning Q2 2024 — the agent presents a structured differential and a flagged list of missing investigations to the triage clinician during the initial patient assessment. The output appears in the department's clinical workstation interface, which at Singapore General Hospital runs on the Sunrise EMR platform. The triage clinician is not required to acknowledge the output before proceeding; it is available as a reference. Where the agent activates a high-acuity flag — sepsis, acute coronary syndrome, or paediatric atypical fever — the output escalates to the senior clinician on duty's screen automatically. This escalation pathway was designed to mirror the workflow already used for the department's manual triage escalation process, so that the agent's alerts arrive in a context clinicians recognise rather than introducing a new cognitive layer. Tan's team ran a three-week simulation across two emergency departments in January 2024 using historical case data to verify that the alert volume was operationally manageable before the live deployment design was finalised.

The second-order signal

Singapore's deployment is not, primarily, a story about diagnostic agents. It is a story about what happens when a small, high-income, linguistically complex health system decides that the governance infrastructure for clinical AI is a sovereign technical problem rather than a capability it can import from a larger jurisdiction. The PRISM framework is a Singapore product. The polylingual consistency threshold is a Singapore requirement. The HPRG's SaMD guidance update — drafted in dialogue with the cluster's evaluation team — is a Singapore regulatory instrument. The procurement review precedent establishing language performance as a legitimate technical criterion is a Singapore legal outcome. None of these would have existed if the cluster had followed the path of least resistance and accepted an international vendor's FDA-cleared portfolio as the evaluation standard.

The second-order effect is already moving. Malaysia's Ministry of Health convened an inter-agency working group on clinical AI governance in January 2024 that is understood to be reviewing the PRISM framework's gate structure as a reference point for Malaysian SaMD guidance under development. Indonesia's Badan Pengawas Obat dan Makanan — the food and drug authority — has submitted a formal information request to Singapore's HPRG covering the November 2023 guidance update. Thailand's National Health Security Office is in early dialogue with MOH Holdings about a technology cooperation arrangement that would include the diagnostic agent architecture. The framework Singapore built to solve its own multi-language clinical problem is becoming the governance template for a region of 680 million people whose clinical AI regulatory infrastructure was, eighteen months ago, a blank page.

Tan's team will not have published PRISM publicly by the time this briefing goes to press. The framework circulates through the clinical informatics community in the Asia-Pacific region the way Mayo's CARS framework circulated through the American academic medical centre network — conference presentations, bilateral Ministry exchanges, and the kind of structured information sharing that happens between Chief Medical Information Officers who are solving identical problems at a six-month offset from each other. The institutions that move on governance first will not simply face a cleaner regulatory pathway. They will have set the terms on which the rest of the region is evaluated.