• SAS says healthcare AI still faces evidence and data gaps.
• SAS says clinical AI still needs human oversight.

Artificial intelligence is being used across healthcare and life sciences, but clinical applications still face questions around evidence, data quality, bias, and accountability.

In an exclusive interview with Tech Wire Asia, Mark Lambredt, Senior Director for Global Health and Life Sciences at SAS, discussed where AI is being applied and what healthcare organisations need to address before using it in patient care.

“AI can never replace that kind of decision, and it should not replace that kind of decision,” Lambredt said. “It should be an input to the doctor making the decision.”

Clinical AI systems need to produce results that are repeatable, traceable, and supported by evidence. Clinicians may not understand every technical component behind a system, but they need enough information to assess why it reached a recommendation and whether its output can be trusted.

“Enabling them to trust it means we need to have evidence that it works,” Lambredt said.

Clinical validation remains part of that requirement. Lambredt pointed to computer vision as one area where algorithms are being used despite what he described as a lack of clinical trials establishing whether some systems work as intended.

Evidence and data remain central to healthcare AI

The quality and composition of health data affect whether an AI model can be transferred between patient populations.

Lambredt pointed to work in southern Denmark involving laboratory data from liquid biopsies, including blood and urine tests. Automated laboratory processes and analytics are used to calculate patient risk and flag cases that may require further investigation, including potential oncology risks.

The system is intended to identify patients who may warrant closer examination rather than make the clinical decision itself. Applying the same algorithm elsewhere, however, requires further testing.

Population characteristics and genetic predispositions can differ between countries. Lambredt used Denmark and Singapore as an example, arguing that an algorithm developed using one population cannot simply be transferred to another with an assumption that it will perform identically.

Organisations therefore need to test models against the populations in which they will be used, identify potential bias, and examine whether particular variables have an outsized influence on the results.

AI applications already span drug discovery, clinical development, manufacturing, supply chains, and healthcare delivery, although their use differs considerably by application.

Drug discovery is one area where Lambredt sees established uses before medicines reach human testing. AI is also being applied in clinical development to support trial protocol design and patient selection.

He was more cautious about using AI to replace clinical trials altogether, particularly proposals to simulate trials without testing medicines in patients.

“We’re nowhere near there,” Lambredt said. “Biology is so complex, and we need to test drugs in patients.”

Clinical healthcare presents another challenge. Lambredt put the proportion of healthcare AI being developed that reaches hospital practice at 2%, attributing the gap largely to insufficient evidence.

Retinal imaging is one example he cited of the divide between demonstrated capabilities and routine clinical use. Research has explored whether retinal scans can help predict conditions beyond eye disease, but Lambredt said such capabilities are not necessarily being used in practice.

Asked about the pace of AI development in healthcare, he said progress remains constrained by the fragmented structure of the sector.

“It is improving, but probably too slowly,” Lambredt said. “Because of the fragmented nature of healthcare. Let’s not forget that, to apply AI, you need to have good data.”

Electronic health records provide part of the information required for predictive systems, but other sources can include laboratory and demographic data. Those datasets can sit across separate systems and organisations.

Bringing them together introduces privacy and governance requirements alongside the technical work. Health information needs to be protected against inappropriate use while rules are established for legitimate secondary uses, including research.

Lambredt argued that patients should retain ownership of their data while approved secondary use should remain possible for research into new therapies.

AI needs to fit clinical and operational workflows

Clinician training is another requirement. AI tools and their applications continue to change, making one-off training insufficient for healthcare teams expected to work with them.

Healthcare organisations also need to consider how AI fits into existing clinical workflows. A system can provide an accurate result but still offer limited practical value if that information arrives too late to help with a decision.

Lambredt used medical imaging as an example. If a doctor can identify a condition from a CT scan immediately, waiting considerably longer for an AI system to return the same finding does not improve that part of the workflow.

“You can apply AI blindly, but will it really help a doctor?” Lambredt said.

Operational processes provide a clearer area for automation than decisions directly affecting patient care.

Prior authorisation, financial processes, scheduling, and patient flow are among the areas Lambredt identified for greater automation. AI can also support the coordination of staff and resources around emergency department demand.

Clinical decisions require a different approach. AI can provide information to support an assessment, but the physician remains responsible for deciding how a patient is treated.

Generative AI can also assist with information-heavy tasks. Summarising a patient’s medical history before a consultation, for example, can reduce the amount of documentation a clinician has to review manually.

The risk of inaccurate or fabricated outputs means those summaries still require oversight.

“You need the human in the loop,” Lambredt said. “Actually, not just in the loop — it’s the human and the AI working together to make the right decision.”

AI also has applications beyond individual patient care. Lambredt pointed to population health and disease monitoring, including the use of health data to identify chronic disease risks and determine where preventive programs may be needed.

Singapore is among the health systems he identified as exploring the use of population-level data to assess the risk of chronic diseases developing over a 10- to 20-year period. He also pointed to Finland, the Baltic states, and Japan as examples of countries exploring preventive approaches.

Accountability and regulation shape deployment

Questions around responsibility become more acute when an AI-supported decision contributes to patient harm.

Technology providers need to work with clinicians, healthcare institutions, and health authorities because technical teams do not necessarily have the clinical expertise required to assess every effect an algorithm could have on patients.

“Ultimately, it’s the responsibility of those who develop the algorithm to be responsible for the decision,” Lambredt said.

Generative AI adds another layer because healthcare organisations can use foundation models built by third parties. Users cannot be expected to understand every component of those models, making testing, monitoring, and controls particularly important for higher-risk applications.

Regulators are also setting more specific expectations for AI in medicine development. The US Food and Drug Administration and European Medicines Agency jointly published 10 principles for good AI practice in drug development in January 2026.

The principles cover the medicines lifecycle from early research and clinical trials to manufacturing and safety monitoring. They include a risk-based approach alongside requirements around data governance, model development, performance assessment, and lifecycle management.

For life sciences companies, clearer requirements can provide a framework for determining what evidence and monitoring are needed when incorporating AI into drug development.

“It gives certainty to life sciences companies so they know what they have to respect, what they have to invest in, and what the roadmap needs to look like if they apply those algorithms,” Lambredt said.

The regulatory process also requires input from technology providers, pharmaceutical companies, healthcare organisations, and policymakers. Technical capabilities have to be considered alongside the data available to life sciences companies and the risks associated with each use case.

Healthcare carries additional considerations because AI deployments can involve patient safety, privacy, and the secondary use of health data.

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.

Tech Wire Asia is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.