IPM Take
Most healthcare AI is presented as universal intelligence, but it is not.
Models learn from particular patients, institutions, coding practices and clinical guidelines. When those models are moved across borders, they can carry foreign assumptions into local care.
Singapore is trying to confront that problem directly by adapting foundation models to its own multi-ethnic population and public healthcare system. Starting with diabetes, hypertension and hyperlipidaemia makes sense because these conditions are common, interconnected and managed largely through primary care.
But “local” does not automatically mean fair, safe or clinically useful.
A model trained on Singaporean data can still underperform in minority communities, amplify gaps in access or turn past clinical decisions into future recommendations. The programme should not be judged by whether its AI sounds more Singaporean. It should be judged by whether patients receive earlier, better and more equitable care after clinicians use it.
Executive Summary
Singapore has announced the initial clinical focus of the Singapore Medical Foundation AI Model programme, known as SIMFONI. The government-supported initiative will adapt existing foundation models to Singapore’s population, clinical guidelines and healthcare workflows instead of relying solely on models trained mainly on Western data. It will begin with clinical decision support for cardiometabolic conditions in primary care and multimodal AI for eye diseases.
SIMFONI was established under the Consortium for Clinical Research and Innovation, Singapore in September 2025. It is supported by the Ministry of Health through the National Medical Research Council Office at MOH Holdings. Professor Robert Morris was appointed executive director, with the programme focused on translating AI research into safe and responsible clinical use.
For cardiometabolic care, the proposed decision-support system is expected to use validated risk models, national clinical guidance and local patient data to help primary care clinicians identify risks and consider diagnoses, treatment pathways and next steps. Singapore’s health minister has stressed that the models should support rather than replace clinical decision-making and should eventually be integrated into public healthcare workflows when ready.
The disease burden makes the choice politically relevant. Singapore’s latest national survey found diabetes prevalence of 9.1%, hypertension prevalence of 33.8% and hyperlipidaemia prevalence of 30.5% among residents. Obesity prevalence increased from 10.5% in 2019-2020 to 12.7% in 2023-2024. Approximately one in three residents has hypertension and/or hyperlipidaemia.
Singapore is not beginning from zero. The Ministry of Health has separately announced that its ACE-AI risk tool will begin rolling out to primary care from early 2027. ACE-AI identifies people at high risk of developing diabetes or hyperlipidaemia within three years, with those classified as high risk offered subsidised annual cardiovascular screening rather than screening every three years. SIMFONI appears to represent a broader next step, moving from a specific predictive tool toward adaptable clinical foundation models.
The programme should therefore be treated as a Signal. Its national backing, planned clinical integration and emphasis on local data make it more consequential than another research prototype. But deployment timelines, validation results, patient outcomes and governance arrangements have not yet been published in sufficient detail to judge whether it will improve care.
Why it matters
- Policymakers and public authorities: Singapore could establish a model for nationally governed healthcare AI, but public accountability must cover data use, procurement, safety monitoring and the consequences of algorithmic errors.
- Clinicians and primary care providers: Decision support could reduce cognitive and administrative burden, but poorly designed alerts may create additional work, defensive medicine and uncertainty about who is responsible for acting.
- Regulators: Foundation models can be adapted, updated and applied across multiple tasks. Oversight must address how performance changes after deployment, not only whether the first version passes evaluation.
- Patients and advocates: Using local data may improve relevance, but patients need transparency about how their records are used and whether AI has influenced screening, referral or treatment decisions.
- Data and AI developers: Localisation must be demonstrated through subgroup validation and prospective clinical evidence, not simply through fine-tuning on national datasets.
Healthcare AI has a geography problem.
Many of the most advanced models have been trained in the United States and Europe, using data from large academic hospitals. They learn from those patients, those guidelines and those clinical systems.
Then they are marketed as general-purpose intelligence.
Singapore is challenging that assumption.
Through the publicly funded SIMFONI programme, the country plans to adapt existing healthcare foundation models to its own multi-ethnic population, national guidelines and public-sector workflows. The first priorities are cardiometabolic decision support in primary care and multimodal AI for eye disease.
The logic is straightforward. A model trained mainly on Western data may not understand Singaporean disease patterns, treatment pathways or population diversity well enough to support safe clinical decisions.
AI does not become neutral because it processes more data.
It becomes more confident about whatever the data taught it.
Cardiometabolic disease is an obvious place to test this approach. Diabetes, hypertension and hyperlipidaemia are common, frequently overlap and create sustained pressure across primary care, hospitals and specialist services. A clinician may need to interpret years of blood pressure readings, laboratory results, medications, kidney function and cardiovascular risk data during a short appointment.
A foundation model could bring those fragments together.
It could flag rising cardiovascular risk before symptoms appear. It could identify treatment gaps, suggest guideline-based next steps and help clinicians decide who needs closer follow-up.
But every recommendation contains a policy decision.
What level of risk should trigger an alert? Should the clinician order more tests, change medication or refer the patient? What happens when the algorithm and the doctor disagree?
Once embedded into an electronic record, the model is no longer simply analysing healthcare policy.
It is helping deliver it.
Singapore’s broader AI strategy already offers one encouraging principle: prediction should lead to action.
Its separate ACE-AI initiative is expected to identify people at high risk of developing diabetes or hyperlipidaemia within three years. High-risk individuals will be offered subsidised cardiovascular screening annually rather than every three years.
That matters because a risk score without a response pathway is not preventive care.
It is an electronic warning.
SIMFONI could take this further by helping guide diagnosis, treatment and referral. But foundation models are also harder to govern than conventional risk calculators. A calculator performs one defined task. A foundation model may combine multiple data sources and be adapted for different uses.
It may perform well when detecting uncontrolled hypertension but less reliably when suggesting treatment changes. It may summarise a patient record accurately while producing inconsistent referral advice.
There will not be one number proving that “the model works.”
Each clinical use will need its own evidence.
Using Singaporean data may improve relevance, but localisation does not automatically create fairness.
Health records reflect who accessed care, which tests were ordered, which diagnoses were recorded and whether patients returned for follow-up. If some communities historically received later diagnoses or fewer referrals, a model can learn those patterns as normal practice.
That is the paradox of local AI.
A system can become more representative of local healthcare while also becoming more faithful to its existing inequalities.
Singapore’s ethnic diversity makes subgroup testing essential. Performance should be reported by age, sex, ethnicity, socioeconomic position and multimorbidity. Average accuracy can conceal selective failure.
A model that performs well overall but poorly for a smaller population may still look successful in a national evaluation.
For the patients it misses, the national average offers little protection.
Responsibility is another unresolved issue.
Singapore has emphasised that SIMFONI should support rather than replace clinicians. The doctor will remain the final decision-maker.
In practice, that boundary may blur.
An AI recommendation inside a national electronic medical record carries institutional authority. Clinicians may follow it because they trust the system, lack time to reconstruct its reasoning or fear the consequences of ignoring it.
The clinician may technically make the final decision.
The software may still shape that decision before the consultation begins.
Doctors therefore need to understand why a recommendation appeared, which data were used and how confidently the model reached its conclusion. They must also be able to override it without facing unnecessary administrative pressure.
Accountability cannot be postponed until something goes wrong.
If a patient is harmed after a clinician follows incorrect advice, responsibility could fall on the clinician, hospital, developer, data provider or government programme that approved the system.
Leaving that question unanswered effectively transfers risk to frontline staff.
Foundation models also change over time.
They may be updated with new data, adapted to additional clinical tasks or connected with new information sources. A change intended to improve one function could affect another. Performance may drift as clinical guidelines, coding practices and disease patterns evolve.
SIMFONI will therefore need continuous oversight, not a one-time approval.
Version control, audit trails, change documentation and real-world performance monitoring should be built into the programme from the beginning. Major updates may need renewed clinical validation rather than being treated as ordinary software maintenance.
The national dataset behind the programme will be equally powerful and politically sensitive.
A standardised dataset could reduce fragmented development and make it easier to test models across Singapore’s public healthcare clusters. But centralisation also concentrates responsibility.
Health records can reveal diagnoses, medications, family relationships, genetic risk and patterns of healthcare use. Even de-identified multimodal datasets can carry privacy and re-identification risks.
The policy question is not only whether the data use is lawful.
Patients should know how their information contributes to model development, who can access it and whether publicly funded models or outputs may later be commercialised.
Public funding should produce public value.
A national dataset should not become a subsidised training resource for products the health system later has to buy back.
The strongest feature of Singapore’s approach may be its stated restraint.
Health Minister Ong Ye Kung has warned against deploying AI simply because it can detect more abnormalities. A tool should identify findings only when the health system understands what they mean and has a credible pathway for diagnosis, counselling and treatment.
That principle should define SIMFONI.
Healthcare does not need AI that finds everything.
It needs AI that identifies what clinicians can interpret, explain and act upon.
A technically impressive algorithm may still fail patients if confirmatory testing is unavailable, referrals are delayed or clinicians are overwhelmed by alerts.
The programme should therefore be judged by clinical outcomes rather than processing speed.
Did blood pressure control improve?
Were high-risk patients diagnosed earlier?
Did referrals become more appropriate?
Did clinician workload fall?
Did inequalities narrow?
Were cardiovascular events prevented?
Accuracy is necessary.
It is not the endpoint.
Singapore has the infrastructure to become an important test case for nationally governed healthcare AI. Its integrated public system, digital capacity and government investment could allow it to move beyond the isolated pilots that dominate medical AI elsewhere.
But national scale magnifies both success and failure.
A weak model used in one hospital is a local problem.
A weak model embedded across a national health system becomes infrastructure.
SIMFONI should move carefully enough to build evidence, but not so cautiously that it becomes another permanent pilot. Validation results, subgroup performance, clinician overrides, adverse events and patient outcomes should be reported publicly.
Failures should be published too.
Trustworthy AI is not AI that never makes mistakes.
It is AI governed by institutions willing to identify, disclose and correct them.
Singapore is trying to teach healthcare AI its local medicine.
The harder task will be teaching the health system when not to listen.
