Across healthcare, AI-enabled diagnostic solutions are flooding the market. From cardiology and imaging to oncology and primary care, clinicians are being asked to evaluate algorithms that promise earlier detection, better accuracy, and more efficient care. Many of these tools are technically impressive. Some are FDA-cleared. A few even come with early clinical data. Yet only a relatively small number have become part of routine clinical practice.
That gap is not about fear of AI. It reflects growing realism among clinicians and health system leaders who have learned that technical performance alone does not translate into clinical impact. The real question is no longer whether AI belongs in diagnostics, but how to identify which tools are supported by sufficient clinical evidence, validation, and operational readiness to be used confidently in patient care.
I spend much of my time speaking with clinicians, health system executives, and regulators across specialties about this problem. While the examples vary by disease area, the underlying challenges are remarkably consistent. Cardiology, in particular, offers a clear view into why so many AI solutions struggle to scale and what differentiates those that are poised for adoption.
Moving past the obsession with false positives
One of the most common concerns about AI diagnostics is the risk of overwhelming the system with unnecessary follow-up testing. The assumption is that earlier detection inevitably leads to more false positives and downstream burden.
That concern is understandable, but it is not always supported by evidence.
In a study published in Nature Medicine, U.S. researchers randomized 120 primary care teams (181 clinicians) caring for more than 22,600 patients in a real-world primary care setting and showed that ECG-AI increased the detection of heart failure among asymptomatic patients by 30%, with no statistically significant increase in the number of echo studies ordered. More patients were identified earlier — when heart failure treatment is more effective and less costly — without overwhelming the system.
But the role of AI can extend beyond early detection of known disease to insights that help us identify risk before it becomes clinically obvious. In another study, researchers examined more than 100,000 patients who underwent both ECG-based assessment and echocardiography to evaluate low ejection fraction (LEF). In about 90% of cases, the two tests agreed on classifying patients as normal (79%) or low (11%) ejection fraction.
The more interesting group was the remaining 10% of patients whose ECGs suggested LEF, even though their echocardiograms appeared normal. Rather than writing those cases off, the team reviewed their mortality over time. It turned out that five years later, the ECG-AI-positive, echo-negative group had a 65% higher all-cause mortality than patients who were negative on both tests.
What looks like a false positive in the moment often turns out to be an early signal. AI can often identify risk earlier than clinicians would, even when traditional tests appear normal. These findings suggest that AI is not creating noise or flooding the system; rather, well-validated AI can provide new diagnostic insight, expose gaps in existing clinical pathways, and help clinicians act earlier and more deliberately.
Why so many AI tools never move past the pilots
Before validation even begins, the strength of the underlying data matters. AI models trained on large, diverse datasets — spanning structured data, unstructured clinical notes, and longitudinal patient records — are better positioned to achieve the most robust predictive performance across a variety of patient populations. Without that foundation, even well-designed algorithms risk breaking down outside controlled settings.
Despite the volume of innovation, many AI-enabled diagnostics struggle to move beyond limited pilots, highlighting the importance of clinical validation, workflow fit, and effective execution, meaning how a tool is implemented, supported, governed, and sustained in everyday clinical operations. In practice, the same issues come up again and again: clinical validation, ease of use, workflow integration, and system-level value.
Clinical validation comes first. Regulatory clearance is necessary, but it does not answer the question clinicians care most about — whether this technology meaningfully improves patient care. Confidence is built when evidence extends beyond minimum regulatory thresholds and is anchored in randomized clinical trials conducted in real-world care settings. Other forms of evidence can complement validation over time, but randomized studies remain the strongest signal that a tool improves patient outcomes.
Ease of use is often misunderstood as a question of interface design or technical simplicity. Too many AI tools are built by engineers and used by doctors, rather than by physicians for physicians. In practice, clinicians judge AI tools by how much friction they introduce into everyday care. A solution can perform well in isolation, but if it adds steps, interrupts familiar workflows, or shifts cognitive burden onto already stretched clinical teams, it will face resistance.
Those challenges become most visible at the point of workflow integration, where many promising AI diagnostics struggle. Clinicians across specialties are already managing heavy cognitive and administrative loads. Tools that require extra steps, separate interfaces, or manual data movement rarely last. The AI solutions that gain traction are the ones clinicians barely notice because they fit naturally into existing workflows.
Health systems also place a heavy emphasis on return on investment. Many AI diagnostics, especially those used early in the care pathway, provide value from the moment they are deployed, with additional benefit accruing over time as patients are identified earlier, treated sooner, and managed more effectively. Their impact becomes clearer through downstream changes in referrals, care decisions, and long-term outcomes. Many payers have recognized this and now offer reimbursement for relatively inexpensive tools that enable earlier detection and help avoid costly late-stage interventions. In this way, AI can uniquely deliver both clinical improvement and financial return, moving from simply cost-effective to cost-saving at the system level.
Turning clinical confidence into adoption
Once clinicians are comfortable with a validated software device, the focus shifts beyond the bedside. At that point, the challenge is less about belief and more about execution. The next question becomes whether the organization is positioned to support adoption in everyday practice.
AI is still new enough that most health systems are still defining how AI decisions are made and owned in real time. In many cases, AI technologies follow the same review and procurement processes as general software, even though an increasing number are designed and regulated as software as a medical device (SaMD) and are built to meet the standards required for direct clinical use. As a result, the last mile of adoption remains slow, even as innovation accelerates. AI development cycles are compressing rapidly, with product iteration increasingly measured in months rather than years. That pace increases the burden on health systems, which must evaluate, oversee, and sustain tools that evolve faster than traditional enterprise software. Adoption, however, still depends on clinical change management, IT prioritization, and governance structures that move more deliberately. Closing that gap will be critical as AI innovation continues to outpace traditional deployment models.
This helps explain why adoption varies even among FDA-cleared, software-based medical devices. Clearance establishes safety and performance. Whether a tool is actually used depends on how well it fits existing workflows and organizational priorities. When that fit is right, adoption tends to follow.
Trust, accountability, and what it takes to last
The more friction a new AI diagnostic device introduces across clinical, administrative, and technical workflows, the less likely it is to become part of routine care. Adoption depends on more than usability alone. It requires strong clinical evidence, clear operational pathways, and the ability to deploy technology efficiently across teams.
Healthcare organizations are right to scrutinize AI through the lenses of ethics, data governance, cybersecurity, and compliance. These are table stakes. What matters just as much is whether AI is used deliberately to improve care, not simply to reduce risk. Tools that earn trust demonstrate real clinical value, fit naturally into existing workflows, and support better decisions over time.
Skepticism toward AI is understandable, particularly as the field continues to mature and distinguish tools that withstand rigorous clinical testing from those driven primarily by novelty. If AI is going to earn a lasting role in diagnostics, it must be held to the same standards as any serious medical advance: demonstrate real-world clinical performance at scale, ideally supported by randomized clinical trials and longitudinal evidence tied to meaningful patient outcomes.
How AI will reshape diagnostics
Over the next several years, AI is likely to reshape diagnostic care in ways that are more evolutionary than dramatic. This shift will be gradual and designed to support clinicians, building on their expertise rather than replacing it. What will change is the information that they have access to and the timing.
In cardiology, AI-supported algorithms are already making parts of diagnosis more objective by drawing on large, real-world datasets rather than relying solely on individual pattern recognition, offering both simplicity and valuable insights compared with the best tools previously available to clinicians. That objectivity can reduce cognitive burden and allow clinicians to spend more time on interpretation, communication, and shared decision-making.
From hype to practice
AI in cardiology does not need more enthusiasm. It needs disciplined implementation.
At its best, ECG-AI functions like a routine clinical marker. It adds context early in the visit without disrupting workflow or replacing clinical judgment. It informs decision-making rather than directing it.
The AI software that succeeds will be grounded in strong clinical validation, introduced through workflows clinicians already use, supported by fit-for-purpose oversight structures and implementation, and ultimately shown to improve patient outcomes. When those conditions are met, adoption follows, not because AI is fashionable, but because it works.
Photo: BrianAJackson, Getty Images
Simos holds a Medical Degree from the University of Athens, Greece, an MBA with distinction from Harvard Business School, a Master’s in Public Health from Harvard University, and is a CFA charterholder.
This post appears through the MedCity Influencers program. Anyone can publish their perspective on business and innovation in healthcare on MedCity News through MedCity Influencers. Click here to find out how.