Can AI Clear the Echocardiogram Backlog? NICE's Verdict on Heart Failure Tech

Maya: Imagine taking a complex diagnostic analysis that usually takes a human expert 587 seconds and shrinking it down to exactly 3.2 seconds using artificial intelligence.

Sam: It sounds like a total game changer for overwhelmed cardiology clinics. But today's document asks the real question: does saving a few minutes on a computer actually mean patients get treated faster?

Maya: Welcome to AI in Medicine - Smart Summaries. Just a quick reminder right at the top, our voices are AI-generated, and this is a summary of a public document.

Sam: Today we are unpacking an early-use assessment from the National Institute for Health and Care Excellence, or NICE. It was published on 19 May 2026.

Maya: The document evaluates artificial intelligence-assisted echocardiography analysis and reporting to support the diagnosis and monitoring of heart failure. And their bottom-line recommendation is incredibly important for health-system leaders and tech companies.

Sam: Right. The verdict from NICE is that there is not enough evidence to support funding these AI technologies in the NHS right now. Access should only be through company, research, or non-core NHS funding.

Maya: Exactly. They looked at 4 specific technologies. But before we get into why they pumped the brakes on funding, we need to talk about why these tools exist in the first place. The clinical burden here is massive.

Sam: Massive is the right word. The document states there are 200,000 new diagnoses of heart failure annually in the UK, and 800,000 people with the condition are on GP registers.

Maya: And despite 40% of people having symptoms that could have prompted an earlier assessment, around 80% of heart failure diagnoses in England happen in the hospital. That is a huge failure of early detection.

Sam: So what does a normal diagnosis pathway look like on the ward?

Maya: For both acute and chronic onset, the initial clinical assessment includes a detailed history, a clinical examination, and a blood test for N-terminal pro-B-type natriuretic peptide, or NT-proBNP. If those thresholds are exceeded, the confirmatory diagnosis relies on transthoracic echocardiography, known as TTE.

Sam: And TTE is really the gold standard here, right? The text says it is used in around 87% of diagnoses.

Maya: Yes. In the NHS, a TTE is usually done in secondary care by a specialist cardiac physiologist, a consultant cardiologist, or a cardiology specialist registrar. It detects abnormalities in the chambers and valves, and measures the heart's pumping ability.

Sam: The text also mentions how heart failure is classified by left ventricular ejection fraction, or LVEF, which is measured by this echocardiography.

Maya: Right. Heart failure with preserved ejection fraction is an LVEF of 50% or more. Reduced ejection fraction is 40% or less. And mildly reduced ejection fraction is the intermediate category of 41% to 49%.

Sam: But here is the bottleneck. A TTE procedure typically takes between 45 and 60 minutes. After that, people still need an appointment with a heart failure specialist to review the findings and confirm the diagnosis.

Maya: And that brings us to the unmet need. In June 2025, there was a waiting list of 156,059 people for echocardiography in England. Only half of hospitals are meeting the NICE quality target, which requires 90% of people referred with suspected heart failure to be investigated using echocardiography.

Sam: They are supposed to be seen within 6 weeks, but only about two thirds of referrals meet that standard. So you have people suffering from severe breathlessness, fatigue, swollen ankles, basically unable to work or live normally, just sitting on a massive waiting list.

Maya: Which is exactly why companies are building AI to help. If AI can aid the interpretation and quantification of these images and automate the report generation, maybe you can reduce that procedure time. Less time per scan means more appointments, which means clearing that 156,059 person backlog.

Sam: Okay, so let's look at the 4 technologies NICE assessed. We have EchoConfidence by MyCardium, EchoGo Heart Failure by Ultromics, Ligence Heart by Ligence UAB, and Us2.ai by EKO Pte Ltd.

Maya: A key detail for our industry listeners: all 4 are designed to aid the operator, not replace them. And their regulatory statuses vary. EchoConfidence and Us2.ai have a digital technology assessment criteria, or DTAC, in place. Ligence Heart and EchoGo Heart Failure do not.

Sam: NICE even looked at their sustainability, which is an interesting modern procurement hurdle. MyCardium and EKO Pte Ltd published Carbon Reduction Plans. Ultromics and Ligence did not disclose theirs. But let's get to the clinical evidence. Why didn't NICE greenlight these for funding?

Maya: It all comes down to the quality of the data. The committee reviewed 19 studies in total. Most of the evidence, specifically 11 studies, focused on Us2.ai. There were 3 studies each on EchoConfidence and EchoGo Heart Failure, and 2 on Ligence Heart.

Sam: And what did those studies actually show?

Maya: They showed that diagnostic accuracy was generally good for detecting abnormalities. But the committee heavily criticized the study designs. The evidence had limited generalisability to NHS clinical practice.

Sam: Why? What was wrong with the studies?

Maya: Many of them were retrospective and observational. They were done in single centres with single operators. Most were done outside of the UK. And crucially, they excluded complex cases and poor-quality echocardiography scan images.

Sam: Wait, really? They excluded poor-quality scans? That feels like building a self-driving car but only testing it on empty, perfectly paved roads. Real-world clinical data is messy.

Maya: Exactly. If you only train and test on perfect images, you don't know what the AI will do when a patient is difficult to scan. Because of this, the committee felt the real-world diagnostic performance in the UK was completely uncertain.

Sam: But what about the time savings? We started the episode talking about a drop from 587 seconds down to 3.2 seconds. That has to count for something.

Maya: That data came from the FEATHER study interim analysis for EchoConfidence. It was a retrospective study based in a UK community care setting. It found the AI reduced the mean time for analysis from 553 seconds and 587 seconds for 2 human readers to 3.2 seconds.

Sam: And there was another study for Us2.ai based in Japan, right? Hirata et al. reported an overall time saving of 524 seconds.

Maya: Yes, but the external assessment group noted there was insufficient detail in the studies to be certain about what the time saving actually related to. The clinical experts said it was unclear whether saving 524 seconds would translate into routine NHS practice.

Sam: Because shaving a few minutes off an analysis step doesn't automatically mean you can squeeze an entirely new patient into the daily schedule.

Maya: Exactly. And it gets more complicated. The clinical experts warned that in some instances, using these AI technologies might actually lead to increases in procedure times.

Sam: Quick note before we carry on. This spot is open for a sponsor. If your company builds or sells AI for healthcare and wants to reach the clinicians, health-system leaders and industry teams who listen to this show, the link to our sponsorship page is in the show notes.

Maya: And now, back to the document.

Sam: Wait, how could an AI make it take longer?

Maya: Because healthcare professionals need to check and review the AI findings. If the AI flags something borderline, or if the operator doesn't trust the automated report, they have to intervene. The committee noted that introducing any further delays could cause harm.

Sam: So there is no evidence to suggest that the time saving results in more people being seen or reduced waiting times. That is the crucial missing link for health-system buyers.

Maya: It really is. And it directly impacted the cost-effectiveness evaluation. The external assessment group had to construct a conceptual Markov model with a 1-year time horizon just to guess the economic impact.

Sam: Because the real-world data didn't exist, they had to make a bunch of assumptions. They couldn't even include EchoGo Heart Failure and Ligence Heart in the model because there was no data on TTE appointment time for them.

Maya: Right. For EchoConfidence and Us2.ai, they assumed a standard TTE appointment takes 45 minutes, allowing for 10 appointments per day. In their base case, they modeled a reduced appointment time of 36 minutes, which bumped the capacity up to 12 appointments per day.

Sam: And if you magically increase from 10 to 12 TTE appointments per day, does the AI pay for itself?

Maya: In this conceptual model, yes. The total cost per scan included software, set-up, IT support, and staff costs. It came out to £4.26 for EchoConfidence and £7.70 for Us2.ai.

Sam: Those are very specific numbers. How did the final cost-effectiveness look?

Maya: EchoConfidence was actually cost saving compared with standard care. The cost difference was -£3.14 and the quality-adjusted life year, or QALY, difference was 0.0005. It saved money mainly because the reduced staff time offset the cost per use.

Sam: And what about Us2.ai?

Maya: For Us2.ai, it was more clinically effective but more costly than standard care. The cost difference was £0.92, the QALY difference was 0.0005, and the incremental cost-effectiveness ratio, or ICER, was £1,674 per QALY gained.

Sam: An ICER of £1,674 per QALY is usually considered highly cost-effective by NICE standards. So why the hesitation?

Maya: Because the entire model is a house of cards built on the assumption that a 45 minutes appointment will reliably drop to 36 minutes. The committee stated clearly that the available evidence for reduced procedure time lacked robustness and generalisability. Therefore, the cost-effectiveness estimates were uncertain.

Sam: There was also a really interesting nuance in the sensitivity analysis about standard cardiology clinics versus one-stop diagnostic centres.

Maya: Yes. In the model, the proportion of patients attending a one-stop clinic was 52%. Clinical experts emphasized that time savings behave differently depending on the setting. In a standard clinic, the wait time from the TTE to the specialist review remains unchanged. But in a one-stop clinic, reducing the TTE time reduces the overall time of the entire visit.

Sam: Which just proves how complex health-system deployment is. You can't just drop an algorithm into a hospital and assume maximum efficiency. The workflow dictates the value.

Maya: Exactly. And there's another major concern the committee raised, which is equity and bias. Heart failure is more prevalent in people from some ethnic backgrounds, including South Asian and Black British populations.

Sam: But we don't know if these AI models were trained on diverse datasets that reflect the UK population. The document says there is a lack of consistency in how external validation cohorts were reported.

Maya: The committee warned that using these technologies without external validation in UK or similar populations may limit their suitability and actually pose clinical risks.

Sam: It's the classic AI blind spot. If it wasn't trained on people who look like me, will it miss my diagnosis? We need more transparency around the populations used for validation.

Maya: Another clinical risk mentioned is that some heart failure symptoms may be caused by other conditions like amyloidosis, valve disease, pulmonary hypertension, and pericardial constriction. The clinical experts worried that some AI technologies might miss these nuances.

Sam: Which brings us back to why a clinical diagnosis is made by a specialist looking at the full picture, not just the AI's output. The technology is an adjunct.

Maya: Right. And while the studies didn't report any adverse events, the committee noted that this could just reflect the retrospective nature of the evidence base. There is no direct evidence of harm, but the clinical risk is still uncertain.

Sam: So what exactly does NICE want to see before they change their minds and recommend funding?

Maya: They laid out a very clear research agenda. First, they need to see the real impact on service capacity. They want data on time saved during procedures in actual NHS secondary, community, or primary care settings.

Sam: They also want to know the impact on the number of appointments per day and waiting times, as well as the actual time taken for a human to review the AI findings.

Maya: They need data on reliability and failure rates. Specifically, how does the AI perform with echocardiograms of varying quality? And how does it perform when used by operators with varying levels of experience?

Sam: And they explicitly called out the need for patient selection and cohorts in the training data to reflect the diverse population seen in the NHS.

Maya: There are a few ongoing studies that might help bridge these gaps. Us2.ai is being investigated in 2 randomised controlled trials, TARTAN-HF and SYMPHONY-HF. EchoConfidence is undergoing a double-blind evaluation in the FEATHER study for unselected consecutive patients in community cardiology services. There are also 2 ongoing studies on Ligence Heart.

Sam: But even with those studies, the committee noted they won't address all the identified evidence gaps. It's going to take a lot of rigorous, real-world proof to get these over the line.

Maya: Which makes this a vital lesson for any company building AI for the enterprise or the health system. Stop relying on retrospective, single-center data that strips out all the difficult cases. You have to prove it works in the messy reality of a clinic, and you have to prove it actually changes the operational math.

Sam: Saving 524 seconds on paper doesn't matter if the waiting list stays at 156,059 people.

Maya: Exactly. To quickly recap: NICE evaluated 4 AI technologies for heart failure echocardiography. Despite promising diagnostic accuracy and potential time savings, the lack of robust, real-world NHS evidence means they are not currently recommended for standard funding. More research on workflow impact, waiting times, and diverse validation is required.

Sam: You can read the full early-use assessment for yourself. The source document is linked in our show notes.

Maya: And remember, we are AI voices summarizing a public document, and this podcast is not medical advice. Thanks for listening to AI in Medicine - Smart Summaries.