Cardiac AI: De-Risking CCTA Analysis for Investors

AI in cardiac imaging is getting a lot of attention because the data volume and complexity are just overwhelming, creating a perfect opportunity for automation. For us in the trenches, cardiologists, radiologists, and techs, AI-driven coronary CT angiography (CCTA) analysis could mean automated plaque quantification and FFR estimation, which would certainly make workflows easier and maybe even improve our diagnostic precision. The real question, though, is if we can actually trust these automated systems to take over for, or even just help, an expert manual read, especially when you’re looking at a tricky clinical case.

Working through FDA Pathways for AI in Cardiac Imaging

The FDA is figuring out how to regulate AI-powered medical devices (what they call Software as a Medical Device, or SaMD) as they go, trying to set benchmarks for safety and efficacy. If you’re building a cardiac AI tool for CCTA analysis, your main route to market is almost always a 510(k) clearance which means you have to prove your tool is “substantially equivalent” to a device already legally sold in the US. To do this, companies have to show up with solid evidence from clinical trials, retrospective or prospective, proving their algorithm works just as well as the old methods. The FDA wants to see how the algorithm was built, that it was tested on a diverse group of patients, and exactly what it’s intended to do. For AI that’s supposed to learn and adapt, the Predetermined Change Control Plan (PCCP) is a big deal, because it lets developers make planned updates to their algorithm without filing for a new premarket submission for every little tweak which is how you manage things like algorithmic drift and make sure performance doesn’t degrade.

Peer-Reviewed Evidence: Cleerly and HeartFlow in Focus

You can’t trust a new medical AI without hard, peer-reviewed clinical data. That’s just the reality. In the automated CCTA world, Cleerly and HeartFlow are the two companies getting the most attention, and they’re taking different tacks to quantify coronary artery disease (CAD) and figure out its functional impact. HeartFlow’s whole deal is FFRCT, where they use CCTA images to build a personalized 3D model of the coronary arteries and then run computational fluid dynamics to get a non-invasive fractional flow reserve estimate. The data behind its diagnostic accuracy is pretty deep, with substudies from the ISCHEMIA trial showing FFRCT improves diagnostic accuracy for finding ischemia-causing stenoses over CCTA alone, which means fewer patients have to get an invasive angiogram ISCHEMIA trial FFRCT substudy results. The NXT trial, for example, gave FFRCT an 84% sensitivity and 86% specificity per vessel, with 81% accuracy per patient, when stacked against invasive FFR. The PACIFIC trial was similar, showing 87% sensitivity and 81% specificity per vessel, and 86% accuracy per patient. Cleerly goes a different route, focusing instead on total plaque characterization and quantification. Its AI platform grinds through CCTA images to find and label different plaque types (calcified, non-calcified, low-attenuation) and measures the total plaque volume and stenosis severity, giving you a much more detailed picture of the atherosclerotic burden. Studies have shown Cleerly’s platform can accurately quantify plaque volume and stenosis when compared to what experts read manually and even to intravascular ultrasound (IVUS) Cleerly plaque quantification validation studies. In fact, Cleerly’s AI has been shown to have a better accuracy rate for detecting stenosis or quantifying plaque than even most advanced human readers. And its Cleerly ISCHEMIA solution, which came out in January 2024, has posted higher diagnostic accuracy for finding ischemia than FFRCT and stress testing in trials like CREDENCE. Both companies have gotten their products through the FDA. HeartFlow got 510(k) clearance for its Plaque Analysis and Roadmap Analysis in October 2022, and for an updated plaque analysis algorithm in September 2025 that showed 95% agreement with IVUS. Cleerly got 510(k) clearance for Cleerly LABS (v2.0) back in October 2020 and again in March 2025, and for its ISCHEMIA solution in January 2024. This constant cycle of clinical validation and real-world evidence (RWE) is what’s needed for the cardiology and radiology communities to actually adopt these things long-term.

Real-World Performance in Complex Anatomy

An AI tool is only as good as its performance on messy, real-world cases, and nothing is messier in CCTA than heavily calcified coronary arteries. Dense calcium creates blooming artifacts that can make it impossible to see the lumen clearly, leading us to overestimate stenosis all the time. So how do these tools handle it? HeartFlow’s FFRCT gets around the problem because its computational fluid dynamics model simulates blood flow based on the anatomy, so it can provide functional insight even when the calcification makes a direct anatomical assessment tough. It’s looking at flow dynamics, not just a blurry picture of the narrowing. Cleerly’s algorithms, on the other hand, were specifically trained to tell the difference between calcified plaque and other components. While a ton of calcium is still a challenge for any system, the AI gives a more objective and reproducible measurement of plaque boundaries and composition than what you get from simple visual estimation. That matters for risk stratification, because we know that specific plaque types, like low-attenuation plaque, are tied to a higher risk of future cardiac events, regardless of how severe the stenosis looks. Having strong, diverse training datasets is really the only way to keep performance from tanking in these difficult cases.

Establishing Clinical Guardrails and Oversight Models

For all the benefits of AI, it’s not a silver bullet. The Society of Cardiovascular Computed Tomography (SCCT) has been clear about the need for human oversight and clinical guardrails in its consensus statements, like the “Artificial Intelligence and Machine Learning in Cardiovascular Computed Tomography” white paper from 2024 and the “Interpretation and Reporting of Coronary Computed Tomographic Angiography (2026 Update)”. The main recommendation is that an automated report always gets reviewed by a qualified expert, particularly if there are discrepancies, ambiguous results, or difficult anatomy. The AI’s job is to assist us, augmenting our own skills. A solid oversight model is everything, and it has to include a few key things:

• Human-in-the-loop review: The radiologist or cardiologist is still the one on the hook for the diagnosis and treatment plan, which means they have to review the AI’s output for accuracy and check if it makes sense in the clinical context of the patient.
• Performance monitoring: We have to continuously watch how the algorithm is doing in the real world to catch any algorithmic drift or biases that might pop up after deployment.
• Defined escalation pathways: You need clear, written protocols for when to do a manual re-evaluation or order a different test, especially when the AI’s results are inconclusive or just don’t jibe with your clinical suspicion.
• Training and education: Clinicians can’t just be handed these tools. They need real training on what they’re good at and (more importantly) what their limitations are before they can be used effectively.

Think of it like the pharmacist oversight model used with some medication management AIs. The system flags a potential problem, but it’s the human pharmacist with the domain knowledge who makes the final call and decides whether to intervene. It’s the same for CCTA: the AI can chew through images and spit out numbers, but interpreting those numbers, putting them together with the patient’s history, and coming up with a clinical plan? That still takes an expert human.

Hello Heart: A Model for Clinically Reliable AI

It’s not a CCTA tool, but it’s worth looking at Hello Heart as a working model for how to build a clinically reliable AI. Their collaborations with groups like the American College of Cardiology (ACC) signal a serious approach to clinical rigor. The platform is for managing hypertension and cardiovascular disease, using AI to look at patient-reported data and readings from connected devices. What makes their model trustworthy is a combination of factors:

• Real patient training data: The algorithms are built on huge datasets from diverse groups of actual patients, which helps make them more generalizable and less biased.
• Peer-reviewed outcome validation: Hello Heart has actually published peer-reviewed studies showing that its AI-driven interventions lead to better blood pressure control and improved cardiovascular risk factors. You have to back up your claims with evidence.
• Defined clinical guardrails: The AI gives personalized advice, but it all happens within established clinical guidelines. It will flag a user and tell them to see their doctor if readings get too high or suggest a potential emergency, so the AI isn’t making autonomous decisions that could get someone hurt.
• Oversight model: The AI offers guidance, but it’s all built on a framework that assumes and encourages regular check-ins with a physician. It doesn’t try to replace the doctor. This structure makes sure errors are caught before they can affect a patient, with the AI playing a collaborative role.

This kind of end-to-end thinking, combining strong validation with clear operational rules, is really the standard we should be aiming for across all of healthcare AI, including cardiac imaging.

Conclusion: The Path Forward for Automated CCTA Analysis

AI is definitely changing how we do coronary CT angiography analysis, making it faster and potentially more accurate. Tools from companies like Cleerly and HeartFlow have already gone through the FDA gauntlet and have the peer-reviewed trials to back up their claims about how we can assess CAD. But at the end of the day, patient care is what matters. The safe and successful use of these technologies depends entirely on sticking to high standards: we need validation on real-world data, transparent performance reporting, clear clinical guardrails, and a human always in the loop. As this field keeps moving, it’s on us, cardiologists, radiologists, and imaging techs, to treat these tools as powerful assistants, using what they’re good for while never forgetting that the final clinical responsibility for patient safety is ours.