Trust in healthcare AI: Addressing challenges in the Asia-Pacific region

• Patient and clinician demand for healthcare AI across the Asia-Pacific region (APAC) is high, but supportive infrastructure to ensure safety and accountability is not fully developed

• Developing robust clinical explainability requires revalidating AI models on the populations they will serve since accuracy established on Western datasets does not always transfer

• Countries across APAC are making different bets on regulatory instruments, creating their own unique approaches to shore up trust and scale healthcare AI

The medical technology landscape in the Asia-Pacific (APAC) region is undergoing a significant structural evolution especially with the introduction of healthcare AI. Across APAC, healthcare systems are transitioning away from merely adopting Western medical technology toward becoming exporters of digital health innovation. For example, South Korean companies such as Lunit have developed high volume AI imaging analysis. Among the first global firms to obtain FDA clearance and CE marks, their technologies have been adopted across multiple US healthcare systems and EU radiology centers.^1,2

The APAC region is home to 60% of the world’s population, but accounts for only 22% of global healthcare spending. Health care worker density is very low, averaging 1.6 doctors per 1000 people.³ Compare that to Organisation for Economic Co-operation and Development (OECD) nations, which average 3.7 per 1000. Healthcare AI offers one possible solution to bridge this resource gap, yet only about 30% of healthcare AI proof-of-concepts across the APAC region successfully cross the chasm into live clinical production.^3,4

I argue that the barrier isn’t a lack of engineering talent, computational power, or demand. The barrier is a systemic trust gap.

For hospital CEOs, university board members, policymakers, and patients, trust in healthcare AI cannot be built on a vague corporate promise. It must be established within clinical, regulatory, and legal frameworks that guide where and when healthcare AI is adopted in clinical practice and how it is reimbursed.

Patients across APAC are not waiting for their health systems to modernize; they are already using AI on their own. In Bain & Company’s 2026 survey of 6,300 consumers across nine APAC markets, nearly 70% reported using generative AI to better understand a diagnosis or treatment plan.³ With two-thirds of APAC patients reporting delays in seeing a specialist (47 days on average), AI is arguably the most responsive part of their care journey.⁵

The pull from inside the healthcare delivery system is just as strong.  AI is one of the few levers that can expand capacity faster than training new clinicians, and clinicians in the APAC region are not AI skeptics: 89% of APAC health professionals believe AI and predictive analytics can save lives through earlier intervention.⁵ Yet, a third of doctors in the APAC region report that their organization is not prepared to deploy AI at scale, and 86% of patients say they would be more at ease with AI if they heard about it from their own doctors.^3,5 Demand is evidently high, but an infrastructure of trust is not in place that can meet it.

Trust in healthcare AI cannot exist without clinical explainability. Traditional AI operates as a statistical black box—data goes in, a recommendation comes out, but the intermediate reasoning remains obscured.⁶ Clinicians are rightfully refusing to adopt tools where they cannot trace the provenance of the AI’s data. To build a trusted environment, the models must be able to clearly show how they arrived at a recommendation or decision. There are three dimensions of explainability to consider:^3,6

1. Mechanistic interpretability asks how the model computed its output

2. Source provenance asks which documents support a claim

3. Local validation asks whether the output is right for this population, this setting, and this intended use

As healthcare systems transition from basic automated workflows to AI agents that can autonomously prepare patient summaries or flag real-time care anomalies, the demand for transparency and explainability has shifted from a preference to a necessity. In the most recent 2026 AMA Physician AI Sentiment survey of 1700 American physicians across a variety of fields, 88% of clinicians said it was either “important” or “very important” that AI safety and efficacy is validated by a trusted entity and monitored over time.⁷ A 2025 KHIDI survey of 2100 Korean clinicians found a lack of information and low reliability were the top barriers to AI adoption.⁸

For an enterprise health system or ministry, implementing an AI tool without local data validation is an existential compliance risk. AI models trained on Western datasets frequently display diagnostic bias when deployed among diverse APAC demographics. In dermatology, for example, 79% of public skin cancer image datasets that report a country of origin come exclusively from Europe, North America, and Oceania, and when state-of-the-art dermatology algorithms were tested on a diverse, biopsy-proven image set, their diagnostic accuracy fell by 27–36%.^9,10 The problem extends beyond medical imaging: Polygenic risk scores, genomic tools increasingly used to predict an individual’s disease risk, are roughly half as accurate in East Asian populations as in European ones, because the genetic studies underpinning them draw overwhelmingly on participants of European ancestry.¹¹ True algorithmic trust, therefore, demands localized validation, ensuring that an AI tool understands the specific clinical baselines, co-morbidities, and socio-economic realities of the exact population it serves.

Ensuring that humans, rather than algorithms, dictate the pace, rules, and final decisions in clinical care is the main pillar of responsible AI governance. Designing healthcare AI without “human-in-the-loop” clinical architecture or end-user involvement decreases quality and risks poor uptake. At Singapore General Hospital, an AI-enabled prescription advisory tool implemented in endocrinology clinics saw limited use when clinicians were unfamiliar with how it had been developed and questioned the clinical relevance and transparency of its recommendations.¹² When frontline physicians and nurses feel that AI is a tool designed to audit them rather than augment them, adoption of even the best tools can slow or even stall. The most successful AI deployments position the technology as a highly capable, collaborative peer. The same AMA survey cited above found that 55% of clinicians prefer to be consulted and 30% want responsibility for implementation of AI in clinical practice.⁷

The APAC region presents a highly fragmented legal and regulatory landscape for healthcare AI.¹³ Major markets are building trust through different instruments, and national differences reveal a lot about how each system considers trust before allowing AI to scale:

• South Korea: The Digital Medical Products Act, the world’s first standalone law for digital medical products, gives software-as-a-medical-device its own legal category rather than stretching conventional device law to fit.¹⁴ Additional regulatory pathways, such as the Innovative Health Technology Assessment, have helped dramatically cut the time needed to secure approval and reimbursement.^15,16

• Singapore: The 2026 update to the national AI in Healthcare Guidelines is less about approving products and more about assigning responsibility. Developers own design, evidence, and post-deployment support, while deployers must run their own governance, validation, and monitoring. Generative AI and continuously learning models are in scope.¹⁷

• Japan: Japan’s distinctive bet is that trustworthy AI starts with a lawful data supply. The amended Next Generation Medical Infrastructure Act created a certified pathway for secondary use of pseudonymized medical records, so models can be built and validated on Japanese patients on a national scale.¹⁸

• Australia: Rather than writing a new statute, Australia’s regulator concluded its existing medical-device framework is largely fit for purpose. They are making targeted reforms, such as reclassifying higher-risk AI tools.  The national safety commission’s AI Clinical Use Guide keeps practitioners accountable before, during, and after every use of an AI tool.^19,20

Legal liability in healthcare AI can be either a structural barrier or an adoption accelerator depending on how clear regulations are written and structured. In jurisdictions with ambiguous frameworks, fear of liability acts as a major barrier. When tort law fails to delineate whether liability falls upon the software developer (under product liability), the hospital system (under corporate negligence), or the treating physician (under medical malpractice), risk-averse clinicians and health executives hesitate to adopt automated tools.^21-23 

Conversely, clear liability frameworks can help accelerate medical AI adoption by establishing de facto safe harbors—not formal immunity, but predictable rules about who answers when AI is involved in care.^21,22 Here are some examples:

• Singapore (accelerator): Through the Health Sciences Authority (HSA) AI medical device guidelines and the AI in Healthcare Guidelines (AIHGle 2.0), Singapore clarifies software classification and data governance. Clear regulatory baselines provide predictability, allowing health systems and Ministry of Health officials to deploy clinical decision support tools with structured “human-in-the-loop” accountability.^24,25

• Japan (transitioning accelerator): By integrating AI devices into the Pharmaceuticals and Medical Devices Agency (PMDA) approval pathways alongside clear guidelines asserting that primary diagnostic liability remains with the licensed physician, Japan makes the allocation of responsibility predictable while setting expected standards of care for clinicians.^23,26 The framework is transitioning because that allocation is under active reassessment: Japanese legal scholars note that as AI takes on more autonomous roles, the physician-centered model will need recalibration. Since standalone software still sits largely outside Japan’s Product Liability Act, vendor accountability is the least settled part of the picture.²³

The question for APAC healthcare leaders is not if AI will redefine care delivery, but how to build the governance framework necessary to scale it securely. The institutions that solve the trust equation today will become the foundation for the region’s outpatient and clinical ecosystems over the next decade.

To unpack these frameworks, review localized case studies, and bridge the gap between AI capability and clinical deployment, register for Healthcare Transformers’ upcoming webinar, “APAC leadership perspectives: An expert panel on building trust in healthcare AI.” The panel discussion will be moderated by Dr Julian Sham, Roche APAC Clinical Digital Transformation Chapter Lead, and will feature:

• Prof. Adam Chee, Singapore General Hospital

• Prof. So Hyun Kang, Seoul National University Bundang Hospital, South Korea

• Prof. Robin Mann, Monash Health, Australia

• Dr. Ernest Lo, Roche Information Solutions, US