Where is autoimmune care in AI preventive care?

Despite the rapid expansion of AI-driven preventive care in Asia, autoimmune and rheumatic diseases, which affect around a tenth of the region's population, remain largely excluded from national AI strategies, creating what clinicians warn is a growing source of preventable, irreversible harm.

In Singapore, where there is world-class specialist capability, primary care time constraints and delayed referrals mean autoimmune conditions are often diagnosed years late, a problem that becomes more severe across Southeast Asia, where some regions have fewer than one rheumatologist per million people.

AI investment patterns have reinforced this gap, with more than 70% of regulatory-cleared medical AI tools concentrated in radiology and cardiology, and little attention paid to multisystem diseases that disproportionately affect women and Asian populations.

Dr Anindita Santosa, a private rheumatologist in Singapore working on GP decision-support tools and AI-enabled education platforms, argues that current data, funding, and model design choices risk encoding gender and population bias into the next generation of preventive care.

In a discussion with Mobihealth News, Dr Santosa shared what it would take to build a more inclusive, preventive AI ecosystem that shortens diagnostic delays in autoimmune disease rather than reinforces them.

Q. Can you expound on the current state of autoimmune and rheumatic care in Singapore – and by extension, Southeast Asia?

A. The state of care can best be described as "high capability, high friction" in Singapore, and "critical scarcity" across much of Southeast Asia.

In Singapore, we have world-class specialist expertise and access to advanced therapies. However, there is a problem with capacity and flow. The true bottleneck sits at the front door of the healthcare system – primary care.

Median consultation times in Singapore polyclinics are typically cited at around 6-9 minutes. That is sufficient for managing hypertension or a viral illness, but fundamentally misaligned with the complexity of early autoimmune disease, which often presents as vague, multi-system symptoms.

Autoimmune diseases affect an estimated 10%-12% of the population (roughly 600,000 people in Singapore), but in day-to-day clinics, these patients are statistically drowned out by far more common acute conditions. Pattern recognition becomes extremely difficult under time pressure.

Across Southeast Asia, this friction becomes fragility. Singapore has roughly one rheumatologist per 100,000 people in the public sector. In contrast, countries such as Indonesia and the Philippines often have fewer than one rheumatologist per million people in many regions. The predictable consequence is late presentation; patients are diagnosed only once disability or organ damage is already established.

Q.  How exactly are autoimmune and rheumatic conditions sidelined in current AI priorities? What kinds of gender, age, or population-level blind spots do you think policymakers and developers are currently underestimating?

A. Autoimmune diseases are consistently deprioritised in AI investment because they are computationally inconvenient. 

Most healthcare AI funding flows toward high-volume, single-organ domains with clean labels and immediate endpoints, such as radiology, cardiology, and oncology screening. This is reflected in regulatory approvals: over 70%-75% of FDA-cleared medical AI devices sit in radiology and cardiology, while rheumatology represents a fraction of a percent.

This creates several underestimated blind spots. For instance, around 80% of autoimmune patients are women. Historically, women's pain and fatigue are more likely to be framed as stress or anxiety in clinical notes. If AI models are trained on this historical data without correction, they learn to systematically down-weight early inflammatory signals in young and middle-aged women.

Preventive AI models cannot simply be imported from Western datasets. Approximately 85%-90% of global GWAS [Genome-Wide Association Study] data comes from individuals of European ancestry; East Asians account for only about 6%. Disease expression also differs. Lupus nephritis affects roughly 50%-60% of Asian patients with SLE [Systemic Lupus Erythematosus], compared with around 30% of Caucasian patients. An AI trained predominantly on Western cohorts will inherently underestimate renal risk in Asian populations.

AI systems not only learn biology; they also learn historical practice patterns. Without intentional correction, they risk automating existing inequities.

Q. From your clinical perspective, what risks does this create for patients living with or developing autoimmune and rheumatic diseases? How much pain or damage could patients have avoided if specialist care had been readily accessible to them?

A. The central risk is preventable, irreversible damage. In rheumatology, we often say "time is tissue."

In rheumatoid arthritis, up to 60%-70% of joint erosions can occur within the first three years if the disease is untreated. Missing the diagnosis in the first six to 12 months often means permanent structural damage.

For systemic diseases such as lupus, average diagnostic delays of four to six years globally can mean the difference between preserved kidney function and lifelong dialysis.

Moreover, we end up paying for the most expensive stage of disease – disability, dialysis, hospitalisation, and biologics – because we missed the cheapest intervention: early recognition. The estimated economic burden of immune-mediated diseases in Southeast Asia exceeds $13 billion annually, largely driven by productivity loss in working-age adults.

Earlier specialist-informed pathways would have prevented a substantial proportion of this cumulative harm.

Q. If AI and genomics are going to define the next decade of preventive care in Singapore and across Asia, what would meaningful inclusion of autoimmune disease look like — in funding, research priorities, data strategy or clinical pathways?

A. Meaningful inclusion means reframing autoimmune disease from a niche specialty into a preventable harm pathway.

National initiatives such as SG100K provide an opportunity to intentionally capture autoimmune phenotypes. We need datasets that link longitudinal primary care notes (the soft signals like fatigue, rashes, and evolving pain) with downstream specialist diagnoses. That is how AI learns what lupus looks like years before kidneys fail.

Funding models tend to prioritise acute mortality (heart attacks, cancer). Autoimmune disease may kill less acutely, but it is a leading cause of long-term morbidity, particularly in women. Investment needs to track the burden of living, not just the cause of death.

AI models must be evaluated on equity. If a model performs well for older men but poorly for younger women, it is clinically unsafe – even if it improves aggregate efficiency.

Q. You're already working with GPs using decision-support tools and AI-enabled education platforms. What have you learned from deploying these tools in real clinics (potentials and limitations) when it comes to earlier detection of autoimmune disease?

A. Through my work with AIGP Health and the development of the Anzu platform, the clearest lesson is that cognitive load is the enemy.

GPs working in high-throughput clinics do not need opaque probability scores. They need tools that remove friction.

What works, for example, is automated or AI-assisted history intake that structures key symptom patterns before the consultation begins. When a GP walks in already seeing "6 weeks of morning stiffness >45 minutes plus new rash," the consult shifts from data gathering to clinical reasoning.

A key risk, however, is false reassurance. If an AI tool labels a patient "low risk" based on limited labs while missing the clinical trajectory, it can delay referrals that clinicians would otherwise have made. In autoimmune disease, AI must function as a safety net, not a gatekeeper.

Q. Finally, what practical steps, whether digital tools, training, or shared-care models, would most help GPs recognise autoimmune red flags earlier, and how could AI be designed to truly support that shift rather than reinforce today's diagnostic delays?

A. To meaningfully shift outcomes, we need to empower primary care with a small number of high-value interventions:

A minimum viable work-up for unexplained fatigue or pain: full blood count, urine dipstick, and CRP. A urine dipstick costs cents; missing proteinuria in a young woman with lupus can lead to dialysis.

Rapid-access or virtual consult pathways for high-suspicion cases, instead of six-month waits.

AI systems must treat the patient record as a time series. A rash today and joint pain four months later may not trigger alarm individually, but together they should.

If AI does not shorten the journey from first symptom to correct pathway, it is not solving the real problem that autoimmune patients face.

_

Dr Santosa's responses have been edited for brevity and clarity.