Can AI-Powered Echocardiography Help Screen for Marfan Syndrome?
Marfan syndrome is a genetic connective tissue disorder that can affect the heart, eyes, skeleton, and blood vessels, most seriously through progressive dilation of the aortic root, which can lead to life-threatening aortic dissection if left undetected. Because early diagnosis and monitoring can be lifesaving, we wanted to explore a question that sits at the intersection of cardiology and artificial intelligence: what role can an AI-powered echocardiography tool like AISAP play in identifying and following patients with suspected Marfan syndrome?
To dig into this, we brought together clinical and AI perspectives for an in-depth conversation. Here's what we learned.
Where This Conversation Started
The idea for this discussion didn't come out of nowhere. It followed an interesting real-world moment at a major cardiology conference, where an AISAP exam captured during a routine demonstration showed an aortic morphology that looked strikingly different from what's typically seen in a general clinical population, prompting an informal discussion among the clinicians present about whether the findings were consistent with a Marfan-related presentation.
That moment turned into a broader question worth exploring properly: when an AI-powered echo tool encounters anatomy this unusual, what should happen next? That's the conversation captured in this post.
How Clinicians Screen for Marfan Syndrome
Marfan syndrome is rarely diagnosed from a single test. Instead, clinicians rely on a structured scoring system, sometimes referred to as the systemic score, that weighs a range of physical findings together:
- The ratio between upper and lower body segments
- Overall height and limb proportions
- Joint hypermobility
- Iridodonesis (a subtle trembling of the lens of the eye)
- Spinal curvature patterns, including the so-called "gothic arch" palate
When enough of these features are present, the probability of an underlying pathogenic genetic variant rises significantly, and the patient is typically referred for an echocardiogram to check for aortic root dilation, one of the two hallmark cardiac findings in Marfan syndrome, alongside mitral valve prolapse.
Importantly, a patient can present with a strong "Marfan habitus," the characteristic body type and physical markers, without yet showing measurable aortic dilation. Even in these cases, clinicians generally recommend a baseline echo and ongoing follow-up, since the aorta can dilate over time. In the United States, patients with a confirmed Marfan diagnosis are typically re-imaged every one to two years to monitor for changes.
Certain physically demanding sports have also been associated with revealing Marfan-like phenotypes, and case reports exist of aortic dissection triggered by high-impact athletic activity, underscoring why timely screening matters, particularly for tall, athletic individuals with connective tissue findings.
Why Marfan Syndrome Is a Hard Case for AI Echo Analysis
Marfan syndrome poses a genuinely interesting challenge for an AI system trained primarily on a general hospital population, where the condition is rare. Because the underlying training data contains relatively few Marfan cases, the aortic morphology seen in more severe or atypical presentations can fall well outside the range of anatomy the model has learned to recognize confidently.
Rather than forcing a low or high score onto anatomy that looks fundamentally different from anything in its training distribution, a well-designed system should recognize when a case is genuinely out of distribution and say so, instead of quietly guessing. This is a meaningful design distinction: a model that confidently mis-scores unusual anatomy is far more dangerous than one that flags uncertainty and asks for human review.
There's also a technical nuance worth highlighting: if an imaging view doesn't capture the maximal width of the aorta, for example, if only the proximal ascending aorta is visualized, even a well-functioning system may underestimate the true extent of dilation. This is a limitation of the underlying imaging, not the algorithm itself, but it reinforces why imaging protocol and view quality matter as much as the analysis layer.
A Proposed Path Forward: Recognize, Retry, Flag
One promising approach discussed was a two-stage recognition process. First, the system attempts to identify the standard parasternal long-axis view using its normal frame of reference. If it isn't confident in that identification, it can retry the analysis without relying on its learned prior about "typical" anatomy, effectively asking, "does this still look like a valid cardiac view, even if the proportions are unusual?"
If the system still can't confidently place the anatomy within a known pattern, the recommended behavior isn't to force a diagnosis; it's to flag the case for human review. This mirrors how experienced clinicians handle ambiguous findings themselves: when something doesn't fit a familiar pattern, the right move is to escalate, not guess.
The conversation also surfaced an important framing for how a tool like this should be positioned: rather than serving as a definitive diagnostic instrument for rare connective tissue disorders, an AI echo tool is best understood as a triage and elimination aid, helping rule things out efficiently in the vast majority of routine cases, while reliably recognizing when a case needs a human expert's attention.
Open Questions Worth Continued Research
A few open questions remain active areas of exploration:
- How much can targeted training improve rare-phenotype recognition? Given how uncommon Marfan syndrome is in most clinical populations, it's an open question whether enough representative cases could be gathered to meaningfully improve model performance on this specific phenotype, or whether flagging for human review will remain the more reliable approach.
- What's an acceptable rate of out-of-distribution flags for rare conditions? Flagging too conservatively could create unnecessary workflow burden; flagging too rarely risks missing genuine outliers. Calibrating this threshold thoughtfully, and validating it against real-world outcomes, is an important next step.
- Should all flagged, atypical cases automatically route to a specialist? Building a clear, reliable handoff pathway from "the algorithm is uncertain" to "a human is now looking at this" is arguably as important as the detection itself.
The Takeaway
AI-powered echocardiography tools have real potential to support Marfan syndrome screening, not by replacing clinical judgment, but by efficiently handling the routine cases and reliably recognizing when something falls outside their competence. The most valuable design principle to come out of this conversation was simple: a good AI diagnostic tool should know what it doesn't know, and hand that uncertainty back to a human being rather than paper over it with false confidence.
That principle, humility built into the model's behavior, may end up being just as important to patient safety as raw diagnostic accuracy.
