Singapore’s Health Minister Ong Ye Kung has a useful way of describing what clinical AI needs. A state-of-the-art smart TV, he says, is worthless without an electrical socket to plug into it. Announcing SIMFONI, a program to train clinical AI models on Singapore’s own practice guidelines and patient data, he made a second observation: The models already available had never attended the local medical school.
That socket, in Ong’s framing, has three pins: a strong digital environment, good data, and sound policy. Most conversations dwell on the first two and treat the third as paperwork to clear once the model works. That ordering gets the risk backwards.
Regulation in healthcare isn’t a formality bolted onto a finished product; it’s part of what makes the product trustworthy enough to use on a patient. Get one pin wrong, and it isn’t merely a compliance problem but a clinical one, and every health system in Asia will have to price it within the next few years.
When AI crosses a border
An algorithm that reads scans accurately in the hospital where it was trained performs worse next door, and worse again in another country. One review of 86 deep-learning tools in radiology found 81% lost accuracy on data from outside their training environment.
The practical result: more false alarms, with doctors chasing findings that lead nowhere. This matters because clinicians already ignore much of what their software tells them, even where AI performs well.
Even in areas where AI performs well, this scepticism holds. JAMA Internal Medicine reported that median physician accuracy interpreting ECGs sat at just 54%. A separate pilot study in the Cardiovascular Digital Health Journal put AI-assisted classification accuracy at up to 95%.
Published research shows doctors override automated alerts anywhere from half the time to almost always, so one more unreliable alert won’t last long before it’s ignored too.
Local disease patterns drive this. Take chronic kidney disease as an example: Prevalence among Singapore residents aged 18 to 74 rose from 8.8% to 13.8% between 2019 and 2022, and diabetic nephropathy is now behind close to two thirds of new dialysis cases.
A model trained where kidney failure usually follows high blood pressure has learned to watch blood pressure closely and blood sugar less so. Fed a Singaporean patient’s chart, where diabetes is the more common trigger, it can let early kidney damage pass unflagged for months while it waits for a blood pressure pattern that was never going to be the cause.
What has to be in place before the model
Singapore has spent years building that groundwork: a national electronic health record, secure access to roughly 50 anonymised research datasets, a shared platform for testing clinical AI, and a common medical record due across all three public healthcare clusters by 2028.
Keeping a model useful once live means retraining and rechecking it against new data, a running cost rather than a one-off purchase that health systems must budget as deliberately as medication or staff.
In many national programs, models are bought first and validated later, if at all. A promising tool often underperforms once it meets real local caseloads, and by then it’s too embedded in clinical workflow to easily pull back.
Policy as trust infrastructure
Healthcare regulation is strict because it handles some of the most sensitive data a person has, and every model built on that data depends on patient trust. That trust must be engineered the way clinical safety is engineered into a drug or device: through rules that hold regardless of who built the tool.
Singapore’s approach has moved fast, and it’s a useful reference for the region: regulators, hospitals, and technology providers built the guidelines together, in step with what was being deployed.
In March 2026, the Ministry of Health and the Health Sciences Authority jointly published the updated “AI in Healthcare Guidelines,” assigning clear responsibility across developers, hospitals, and clinicians. HSA has also opened a regulatory sandbox letting lower-risk AI diagnostic tools from public institutions move faster, with a senior clinician signing off on design and output. Singapore’s data protection law shapes the technology itself: it constrains how records can be pooled and de-identified, so hospitals increasingly train models in a federated way, without centralising raw data.
A rulebook fixed once and left alone will be wrong within a year; one built jointly, and revisited on the same cycle as the models it governs, has a chance of staying right.
A rising caseload, and fewer people to carry it
The pressure is demographic, not technological. Singapore’s resident total fertility rate fell to 0.87 in 2025; citizens aged 65 and above now make up 20.7% of the citizen population; and there are just 2.4 working-age citizens for every senior. That local strain sits inside a global one. The World Health Organization projects a worldwide shortfall of 11 million health workers by 2030, and Singapore is racing the same clock as every other health system.
Diabetes, hypertension, and high lipid levels already affect around 1.8 million people, many of whom do not know it, and chronic kidney disease is projected to reach one in four residents by 2035. Chronic illness does not pause while systems are built. Patients missed this year arrive in hospital in five, needing care that costs far more.
This is why SIMFONI starts here: Cardiometabolic conditions such as diabetes and high blood pressure are the root causes of costlier diseases that follow, including kidney failure, stroke, and vision loss. Catching them early avoids much of that cost and suffering, in a population that can no longer assume most patients will be young and recover.
Translating the model to bedside
Clinicians must be in the room while models are designed, not handed the output at deployment, since they’re the ones who decide whether an alert is worth opening.
Regulators should revisit approvals on a regular cycle rather than rely on a single sign-off, which is exactly what a sandbox model is designed to do. Hospitals and universities must keep supplying local data, and infrastructure needs long-term funding, not a one-off budget.
Most importantly, these models are built on patient records, and consent given grudgingly is a foundation that erodes.
The health systems that get this right won’t be the ones with the best algorithms, but the ones that build AI infrastructure long before a model reaches the bedside. That means building the socket, not selling the model. A state-of-the-art smart TV is still only as good as the wall it plugs into.
