AI Deployments in Clinical Practice

The All-Ireland Clinical AI Congress and AI Tech Expo heard from Dr Amrita Kumar, HealthBay, United Arab Emirates, who delivered a fascinating talk titled ‘Lessons learned in AI deployment’. She is a Consultant Radiologist specialising in breast imaging and is also Clinical Lead at Medica UK. Dr Kumar also works in clinics in Dubai, where she was asked to implement mammogram AI into clinics’ breast screening service in the region.

She described three ‘real-world’ deployments of AI and illustrated the disparity between how an algorithm works on paper, versus deployment in a real-world service. “The algorithm is the easy part – deploying it is the hard part,” Dr Kumar told attendees.

“We have spent a decade asking the question, ‘Does AI work?’ The models work, but the hard question, and the one that decides whether a patient actually benefits, is whether you can deploy it safely and govern it.”

Dr Kumar pointed out that Ireland has approximately half the median amount of radiologists compared to our European counterparts, at 6.5 per 100,000 people, and demand for acute out-of-hours imaging is rising sharply. “You cannot simply recruit your way out of that,” she said. “That is why AI matters.”

The issue also becomes primarily a patient safety one, rather than solely a question of efficiency.

Dr Kumar gave the attendees a synopsis of three deployments of different types of AI, one of which is currently in use in Ireland.

From these examples, she outlined some lessons learned to the Congress.

“Lesson one: AI deployment is a service transformation, not a software one,” she said. “Also, technology alone does not reduce the waiting list. You don’t treat AI as a purchase that just sits on a shelf; it is a service transformation and if you treat it like that, it really delivers. Lesson two: Validate locally, because your own evidence will always beat the vendor claims.”

She continued: “Lesson three: Be honest about what the evidence does and does not show… the lesson is not that we failed to measure something; the lesson is to build the measurement from the start, before you deploy, because afterwards is too late. Lesson four: Make the human accountable and design human factors… the AI does not decide the final report. That is deliberate, and it is what regulators expect right now.

“I think lesson five is the most important,” Dr Kumar told the Congress. “AI is a ‘team sport’. Every statistic I have shown to you exists because of a multidisciplinary team. Clinical specialists, radiographers, AI programme managers, data teams, operations, bookings, clinical governance support, all working alongside the vendor… and I say this to the policy-makers in the room: State AI adoption is not a piece of software you buy; it is the organisational ability that you build in.”

The need for appropriate governance in AI was a theme that ran through the Congress.

Dr Kumar told attendees that her advice is to match governance to the risk exposure.

“One size does not fit all, because the risk is generally very different at any stage,” she said.

“I want you to think about what all parts of this room could be doing to get AI deployment right in this shared journey,” Dr Kumar concluded. “To the policymakers: Fund the validation pathways and share standards that let firms validate these tools properly. To the vendors: Bring us evidence, not claims, and help your healthcare providers to validate locally to generate their own evidence.

“To my fellow clinicians: Own the governance and keep clinical judgement at the centre of it. And to the academics: Build the methods for real-world post-market evidence, because that is the evidence that actually governs deployment. As I said, the algorithm is the easy part, but together we can get the deployment right.”

Dr Kumar’s talk was followed by a lively and informative panel discussion. The panel held a practical discussion on the lessons learned from deploying AI in secondary care, with examples from diagnostic imaging, breast screening, prostate cancer pathways, prescribing, and acute hospital systems. The panel explored workflow integration, clinical governance, implementation barriers and the potential for AI to improve earlier diagnosis and patient outcomes.