Artificial intelligence is becoming more deeply embedded across healthcare, supporting tasks ranging from clinical documentation and medical imaging to administrative reporting and operational decision-making. Yet the path from a promising algorithm to routine hospital use remains complex, particularly when healthcare institutions must address clinical safety, workflow integration, cybersecurity, reimbursement and accountability at the same time.
Across Asia, hospitals are also considering how locally governed AI infrastructure, representative datasets and domain-specific implementation can support wider adoption while protecting sensitive health information. Associate Professor Daniel Ting, Founding Co-Director of the SingHealth Duke-NUS AI Medicine Institute and Chief Data and Digital Officer, discusses where clinical AI is already creating practical value, the barriers holding back routine deployment and the safeguards required to build trust among clinicians and patients.
Where is AI already delivering the most practical value in clinical care today?
Ambient clinical AI is beginning to deliver immediate practical value. It can automatically document consultations, summarise clinical discussions and prepare draft clinical notes, referral letters and discharge summaries. By reducing time spent on documentation and data entry, ambient AI allows clinicians to focus more closely on patients and may help reduce administrative burden and professional burnout.
Medical imaging remains one of the most mature areas of clinical AI. AI is already being used for image segmentation, classification, triage and disease detection in areas such as diabetic retinopathy screening, chest radiography and mammography. Its value is particularly clear in high-volume settings, where it can prioritise urgent cases, support more consistent interpretation and extend specialist expertise to underserved populations.
AI is also improving administrative and operational efficiency. It can assist with meeting summaries, presentations, agendas, administrative briefs and routine reporting. However, the greatest value is achieved when these systems are developed with deep domain expertise. Clinicians, healthcare operators, AI scientists and engineers must work together to ensure that AI addresses genuine clinical problems rather than simply introducing new technology.
What are the biggest barriers to moving healthcare AI from pilots to routine hospital deployment?
The first major barrier is the lack of hospital-grade platforms for real-world deployment. Many algorithms perform well in research environments but remain disconnected from electronic medical records, imaging systems and clinical workflows. Hospitals need secure and interoperable platforms that can deploy multiple algorithms, manage data flows, monitor performance and deliver AI outputs at the appropriate point of care.
The second barrier is insufficient deep domain and implementation expertise. Healthcare AI cannot be developed by engineers alone. Successful deployment requires clinicians, informaticians, software engineers, implementation scientists, administrators, regulators and patients to work together. Without a detailed understanding of the clinical workflow, even a highly accurate model may create additional steps, alerts or risks rather than improve care.
The third barrier is the absence of sustainable reimbursement and business models. Many pilots are supported by short-term grants, but routine implementation requires long-term funding for licensing, infrastructure, cybersecurity, maintenance, validation and governance. AI will only scale when reimbursement models and health-economic evidence demonstrate that it improves outcomes, increases productivity, expands access or reduces overall healthcare costs.
How should hospitals assess whether an AI solution is clinically ready, safe and useful for real-world workflows?
Hospitals should first assess safety and accuracy using a structured framework. In our recently published SAFER framework, safety includes risks such as hallucinations, inappropriate recommendations, automation bias and potential patient harm. Accuracy should be assessed quantitatively through validated performance metrics and qualitatively by determining whether the outputs are clinically meaningful and appropriate for the intended use.
Fairness, ethics and generalisability must also be examined carefully. Hospitals should understand the populations, institutions and datasets used to train and validate the model. Performance should be evaluated across age groups, ethnicities, languages, socioeconomic backgrounds and disease severities. Independent local validation is essential because strong performance in one population or hospital may not translate directly to another.
Hospitals should also evaluate ease of use, regulatory readiness and return on investment. The AI solution must fit naturally within existing workflows and have clear accountability for decisions. Regulatory requirements, cybersecurity, monitoring, maintenance costs and measurable clinical or operational benefits should all be considered before deployment.
Sovereign AI infrastructure is increasingly important. Sensitive healthcare data should be protected through secure, privacy-preserving and locally governed systems. Sovereign AI allows healthcare systems to retain control over their data, comply with local regulations and develop models that are better aligned with local populations, languages, disease patterns and clinical practices.
What lessons from ophthalmology and imaging AI can be applied to other clinical specialties?
One important lesson is that scalable AI requires both deployment platforms and deep clinical domain expertise. Developing an accurate algorithm is only the beginning. Successful implementation requires platforms that connect clinical systems, devices, data and AI models, together with multidisciplinary teams that understand the clinical problem and can redesign the workflow appropriately.
The second lesson is that the mode of deployment should be determined by the clinical context. AI can be used as a first reader, a second reader supporting a clinician, or autonomously for carefully defined tasks. The appropriate model depends on disease prevalence, workforce availability, algorithm performance, clinical risk and the consequences of false-positive and false-negative results.
The third lesson is that health-economic evaluation and reimbursement planning should begin early. Clinical accuracy alone does not determine whether an AI system should be adopted. Hospitals must evaluate whether AI saves time, reduces unnecessary referrals, prevents complications, improves access or allows scarce specialists to focus on more complex patients.
The final lesson is that reimbursement must align with the intended deployment model. AI-first, AI-second and autonomous AI workflows create different costs, responsibilities and benefits. Reimbursement models should therefore reflect the clinical service being delivered and provide sustainable incentives for hospitals and clinicians to adopt effective AI systems.
How can hospitals ensure AI reduces clinician workload instead of adding complexity?
AI should be introduced to solve a clearly defined clinical or operational problem. Hospitals should first identify tasks that are repetitive, time-consuming or prone to delay, and then determine whether AI can simplify or automate them. A successful system should remove steps, reduce duplication and deliver useful information directly within the clinician’s normal working environment.
Clinicians and other domain experts must be involved throughout the development and implementation process. Deep domain expertise helps ensure that AI outputs are relevant, actionable and delivered at the correct point in the workflow. Without clinician involvement, AI may generate unnecessary alerts, duplicate existing work or increase cognitive burden.
Hospitals also require a robust governance framework for safe and responsible AI use. Governance should define who is accountable for reviewing outputs, responding to errors, monitoring performance and escalating safety concerns. It should also establish clear criteria for when AI may be used, when human review is mandatory and when a system should be suspended.
The impact on workload should be measured directly. Hospitals should monitor time saved, documentation burden, alert frequency, clinician satisfaction and patient outcomes before and after implementation. AI systems that do not demonstrate meaningful benefits should be redesigned or withdrawn rather than maintained simply because they are technologically impressive.
What safeguards are needed to build trust among clinicians and patients when AI is used in care delivery?
Transparency and explainability are fundamental. Clinicians and patients should understand what the AI system is intended to do, what data it uses and where its limitations lie. Hospitals should clearly communicate whether AI is assisting a clinician or making an autonomous decision, and responsibility for the final clinical decision must remain clearly defined.
Hospitals should require transparency about the datasets used for training and validation. This includes the size, quality, demographic composition, disease spectrum and clinical settings represented in the data. Any underrepresented populations or known limitations should be disclosed, with additional testing undertaken before deployment in those groups.
Independent and representative validation is essential. An AI system should not be accepted solely on the basis of results reported by its developer. It should be tested on external datasets and, where possible, prospectively evaluated in the local healthcare environment. Post-deployment monitoring is equally important because performance may change as patient populations, workflows and data systems evolve.
Sovereign AI and responsible data governance are also central to trust. Patients and clinicians need confidence that healthcare data remain secure, are governed under local laws and are used responsibly. Local control over data, models and infrastructure can strengthen privacy, accountability and public confidence while reducing dependence on external systems that may not reflect local priorities.
How do you see clinical AI evolving across Asia’s healthcare systems over the next three to five years?
Clinical AI adoption across Asia is likely to accelerate rapidly. Frontier AI and foundation models are already showing increasingly strong performance in interpreting images, clinical text and multimodal healthcare data. Over the next three to five years, AI will move beyond individual algorithms toward systems that integrate imaging, electronic health records, laboratory results, speech and other clinical information.
The greatest value will come from domain-specific and clinically integrated AI. General-purpose frontier models may provide strong foundational capabilities, but healthcare systems will still require clinicians and other domain experts to adapt, validate and govern these models for specific clinical use cases. Deep domain expertise will remain essential for ensuring that AI is safe, relevant and useful in real-world care.
Sovereign healthcare AI will become a strategic priority across Asia. Countries will increasingly seek to develop or host models within trusted local infrastructure, using representative national or regional data. This will help protect sensitive health information, support local languages and disease patterns, comply with national regulations and reduce reliance on external technology providers.
Adoption will vary considerably between healthcare systems. Large academic centres may develop sophisticated multimodal foundation models, while lower-resource settings may prioritise AI for screening, triage and extending specialist access. Each system will require locally validated solutions that reflect its own infrastructure, workforce, population and care-delivery model.
The principal challenge will shift from algorithm development to sustainable implementation. The question will no longer be whether AI can match clinical performance, but whether it can be integrated safely, governed responsibly and financed sustainably. Reimbursement models, health-economic evidence, regulatory clarity and workforce transformation will determine which AI systems progress from pilots to routine care.
