Wearables, ambient monitoring and the case for explainable AI: a conversation with Dr Amel Havkic

Dr Amel Havkic is a pulmonologist, clinical risk manager and Founder of EvoMed Consulting, where his team has built StarMap, an explainable AI platform for clinical decision support. HealthTechAsia’s Matt Brady spoke with him on a podcast hosted by EvoMed Consulting, covering consumer wearables, ambient monitoring, health misinformation and what genuine AI governance looks like in practice.

On whether consumer wearables are gimmicks or genuine diagnostic tools

Brady opened by asking whether smartwatches and wearables are actually effective at measuring things like blood pressure, or whether they are still largely gimmicks.

“If you’d asked me that a few years ago, I would have said this is just a gimmick – you can see your pulse rate, but can you count on what it’s spitting out? I would say no,” Havkic said. “But I do think the algorithms are getting better and better. People are working on them every day.”

He pointed to regulatory approval as the real marker of progress: some devices now carry FDA clearance in their own right. “You have smart devices which have FDA approval themselves, so you’re basically allowed to use them to base a diagnosis on. The most famous example would be the Apple Watch, with its AFib detection.”

In his own practice, he asks patients with symptoms like heart racing whether they own a wearable he can review. Before that was possible, catching an intermittent arrhythmia meant escalating from a 24-hour ECG to a multi-day ECG, and in persistent cases, implanting an event recorder that can remain in a patient for up to five years. He recalled a colleague who had one implanted, found nothing for years, then went into cardiac arrest and had to be resuscitated in the week before the device was due to be removed.

The lesson he draws from that case is that people generally tolerate wearing a watch, but resist more invasive or unfamiliar devices – even ones with real clinical promise. He cited the Apple Vision Pro as an example some in the medtech space believed would reshape operating rooms, but which has not achieved wide clinical adoption. “If it doesn’t feel natural, we don’t want to use it,” he said.

He also drew a regulatory distinction between wellness devices and medical devices: platforms such as Whoop stay positioned on the wellness side, while others, such as Medtronic’s smart bands, take on formal responsibility as remote monitoring devices despite measuring similar data.

On who actually benefits from wearables – and the shift to ambient monitoring

Brady asked whether wearers understand the limitations of their devices, or assume a watch on the wrist means they no longer need to see a doctor.

Havkic said the pattern differs by generation. “There is this ‘let’s generate data with my health’ generation, which is maybe our age, and the younger generation. But the people most in need of monitoring are probably a bit older, and for them that feels unnatural – it’s not something they want to use.” He referenced a recent conversation with a venture capitalist raising the same issue with a wearable-based care model: patients simply weren’t wearing the device enough for it to work.

That gap, he said, is pushing innovation toward ambient monitoring rather than devices worn on the body – citing, as one public example, a bed-adjacent light fitted with motion detectors that can identify if a patient has fallen. “I think instead of thinking only variables, we will get more and more into that ambient monitoring type of scenario,” he said. “It will feel more natural to us than a wearable, even a big one.”

The conversation also touched on the more unsettling end of remote physiological sensing: Havkic referenced reports of long-range cardiac-signature detection technology originally developed for military use. HealthTechAsia has not independently verified the specifics of this claim, but it illustrates a broader point he returned to repeatedly – that as monitoring becomes more invasive and more depended upon, the consequences of that data being manipulated by bad actors grow accordingly.

On resilience: why hospitals still need paper

As a clinical risk manager currently overseeing full ICU digitalisation at his hospital, Havkic argues that resilience has to be designed in from the outset – including for the possibility that the technology fails entirely.

“What do you do if you have a power outage? What do you do if you don’t have any more energy available and the patient treatment is critical?” he said. “The answer is as simple as you could think: paper.” His hospital’s system is built to automatically print full documentation in the background while running on emergency power, precisely so care is not interrupted if digital systems go down.

He also referenced a recent conversation with a cybersecurity specialist about how any remote monitoring system that clinicians rely on to raise alarms is also a system that a malicious actor could manipulate. “The more invasive the technology, and the more the patient specifically depends on that technology, the bigger the potential for doing harm.”

On the EU AI Act and building explainable AI

Asked what has recently caught his attention in health AI governance, Havkic pointed to the EU AI Act’s risk-based approach, which he described as evaluating how much influence a given AI output has over a decision, and how dangerous a wrong decision would be.

He said this thinking shaped the design of his own platform. StarMap can carry out a range of tasks autonomously, but “the whole system is rigged in such a way that no report, and nothing you can draw a conclusion from, is presented to the user directly without one of our experts actively approving it.” Every output includes an explanation of why the system reached that conclusion – an approach he believes should become standard in healthcare AI generally, rather than “a black box which gets a decision out, or a recommendation out, and we don’t know why.”

His internal test for whether a feature needs that kind of guardrail is simple: “Can this harm someone? And if the possibility is that it can, then think about how to mitigate the risk.”

He also referenced a comparison in which a specialised healthcare-focused language model performed on par with general frontier models from OpenAI and Anthropic on medical tasks – his point being that governance and safe deployment matter more than claims of proprietary advantage, since capability gaps between specialised and general-purpose models can close quickly.

On misinformation, overdiagnosis and critical thinking

The conversation turned to the spread of health misinformation, including the trend of diagnosing personality traits or conditions from social media content. Havkic pointed to overdiagnosis – both self-diagnosis via social media trends and, separately, clinically valid but statistically unusual results being misread out of context – as a growing problem he is seeing directly in practice, alongside the more familiar risk of outright false information spreading online.

“The actual problem is the technology is moving way faster than society can adapt, and we’re already seeing the results of that,” he said.

On solutions, he argued that media literacy needs to be taught explicitly, drawing on his own schooling in Bosnia after the Yugoslav war, where pupils in neighbouring classrooms were taught the same history from Bosnian and Serbian textbooks that described it differently. “I don’t believe any information – I don’t even believe information that I give myself. I always keep back some percentage that I might not be right. Critical thinking is something we should learn how to do as a society, and not accept everything just because it comes from a point of authority.”

On the risk of runaway AI capability

Asked how close he thinks the world is to AI systems causing serious, uncontrolled harm, Havkic was direct: “We are playing a dangerous game right now. There’s some type of arms race when it comes to AI capability, and we are one failed lab experiment away from a very bad scenario.”

His recommendation is not to halt development, but to build in the same discipline he applies at EvoMed Consulting: human review before autonomous output reaches an end user, and an explicit account of why a system reached the conclusion it did, particularly in any application where a wrong decision could put a patient at risk.