What AI Is Actually Doing in US Cardiology Clinics Right Now (And What’s Still Broken)

Let’s start with what changed the conversation.

Atrial fibrillation affects 52 million people globally, making it the most common cardiac arrhythmia in the world. In the US, it is a leading cause of stroke, and a large portion of AF-related strokes happen in people who had no known AF diagnosis before the stroke.

The reason? 

AF comes and goes. A standard 12-lead ECG captures maybe 10 seconds of heart activity. If the arrhythmia is not happening at exactly that moment, the test shows nothing wrong.

Once flagged, a patient can be referred for extended monitoring and started on anticoagulation therapy that can prevent a stroke. That is a real clinical outcome, not a marketing claim.

The catch? 

AI-ECG alerts alone do not automatically lead to better outcomes. A pragmatic trial found that while AI-ECG alerts increased oral anticoagulant prescriptions, no reduction in stroke or cardiovascular death was observed in the short-term follow-up window. 

The algorithm’s job is to find the problem. The care pathway still needs humans to act on it, and when that problem is heart failure, the stakes for getting that handoff right go up considerably. 

Heart Failure Readmissions: Where the Financial and Clinical Problem Meet

Heart failure is one of the most frequent discharge diagnoses in the US. The 30-day readmission rate is around 20%, climbing to 30% at 90 days, tracked in this AHA study. Medicare penalizes hospitals for excess readmissions through the Hospital Readmissions Reduction Program (HRRP), which means every preventable readmission is both a patient safety failure and a financial hit.

Traditional readmission prediction has relied on demographic data, primary diagnosis, and comorbidity scores. These models are reasonably good at identifying who is generally high-risk. They are not as good at telling a care team when a specific patient’s condition is about to destabilize.

A meta-analysis covering 943,941 patients concluded that ML models offer promising improvements over conventional scoring systems for predicting HF mortality and readmission, though real-world deployment remains limited.

That last part matters. The gap between a model that performs well in a research dataset and one that actually changes clinical workflows is significant. Hospitals that have implemented AI-driven heart failure monitoring in ways that connect directly to nurse alerts and care coordinator workflows report more consistent results than those where the AI output sits in a separate dashboard that no one has time to check.

LINK-HF2, testing an AI-guided intervention using noninvasive monitoring data for HF patients, found during its pilot phase that workflow integration was the central challenge, not the algorithm itself. 

Clinicians needed the AI’s output to arrive in the right place, at the right time, in a format that supported action. That challenge is worth keeping in mind as the number of tools entering cardiology continues to grow. 

Over 200 FDA-Cleared AI Tools in Cardiology: What That Number  Means

As of early 2026, the FDA device database includes more than 200 cleared algorithms specifically for cardiovascular applications. That is a number worth pausing on.

To put it in context, cardiology and radiology are the two specialties driving the majority of FDA AI clearances in the US. 

A few specific tools worth knowing about:

• HeartFlow Analysis uses CT angiography data to simulate blood flow through coronary arteries and calculate fractional flow reserve (FFR-CT) without an invasive catheterization. In October 2024, the AMA issued a new CPT code for AI-enabled plaque quantification, set to take effect in January 2026, and HeartFlow earned Medicare coverage from 5 Medicare Administrative Contractors. This is meaningful because reimbursement has historically been a major barrier to AI adoption in practice.
• AliveCor Kardia 12L is an FDA-cleared handheld 12-lead ECG system. As of January 2026, its AI (KAI 12L) was FDA-cleared for 39 determinations, including MI detection. The device weighs 0.3 pounds. Trained on 1.75 million ECGs, it brings hospital-grade diagnostic capability to urgent care, primary care, and ambulance settings.
• EchoGo Heart Failure 2.0 from Ultromics analyzes echocardiograms with AI to provide automated measurements of cardiac structure and function, supporting heart failure assessment with greater consistency than manual reading.

The important caveat from AHA guidance issued in November 2025: hundreds of healthcare AI tools have been cleared by the FDA, yet only a fraction have been rigorously evaluated for clinical impact, fairness, or bias. 

FDA clearance confirms a device is safe and substantially equivalent to a predicate. But it does not confirm that the device improves outcomes in your patient population. And for some populations, the gap between what a tool promises and what it delivers is not random.

The Bias Problem Nobody Wants to Talk About in Cardiology AI

Here is the part that is not in most vendor brochures.

AI models in cardiology are trained on data. If that data overrepresents one demographic group, the model learns patterns that perform well for that group and less well for everyone else.

Explainable AI (XAI) approaches, which use methods like SHAP values to show which features drove a model’s output, are emerging as a partial solution. They allow clinicians to see why an AI flagged a patient, which makes it easier to catch cases where the model may be extrapolating beyond its training data.

But even when the algorithm is working exactly as intended, there is a separate problem eating into cardiology, one that has less to do with what AI detects and more to do with everything that happens after the patient leaves the exam room. 

The EHR Problem AI Can Help Solve (If Deployed Correctly)

Cardiologist burnout is not primarily caused by difficult clinical decisions. It is caused by paperwork.

A time-motion study observing 57 US physicians across four specialties found that physicians spent approximately 45% of their total working time on EHR-related tasks.

The failure mode here is worth naming. Deploying AI documentation tools without changing the underlying workflow does not help. If the ambient scribe generates a note that still requires 15 manual corrections before it can be signed, or if the alert filtering system floods a cardiologist’s inbox with anything the algorithm is uncertain about, the tool adds friction rather than removing it.

And some limitations go deeper than workflow. There are things AI in cardiology still genuinely cannot do, regardless of how well it is implemented. 

What AI Still Cannot Do in Cardiology

AI in cardiology is not a replacement for clinical judgment. It is a layer of pattern recognition that operates faster and at higher scale than a human can.

There are specific things it does not do well yet:

• Cross-modal reasoning. An AI model trained on ECGs cannot automatically incorporate a patient’s latest lab values, their social history, and their current medication list into its output unless those data sources are explicitly integrated. Most deployed AI tools in cardiology are single-modal. They look at one data type. The cardiologist still has to synthesize across sources.
• Rare presentations. AI models learn from large datasets. Rare conditions, unusual presentations, and atypical demographics are underrepresented in training data almost by definition. A patient with an unusual form of hypertrophic cardiomyopathy presenting in an atypical pattern is more likely to be missed or misclassified by an AI than a textbook STEMI.
• Explaining uncertainty. Most current AI tools give a prediction or a flag. Fewer of them communicate how confident they are, or what factors would change the output. Clinicians need to know not just what the AI thinks but how much weight to put on it.
• Long-term outcome accountability. A model can predict 30-day readmission risk. Whether acting on that prediction actually reduced readmissions in a given health system requires prospective tracking that most organizations do not yet have in place. None of that means standing still, it means being deliberate about where and how you deploy. 

What a Cardiology Practice or Health System Should Do With This

If you are a cardiologist, practice administrator, or health system leader trying to make sense of where to invest:

• Start with the problem, not the tool. Which patients in your population are you most likely to miss? High-volume, low-complexity cases benefit most from AI-assisted screening. AI adds the least value where experienced clinical judgment is the dominant input.
  
• Require subgroup performance data in writing before signing. Verbal assurances from a vendor do not hold up. If the contract does not include demographic performance benchmarks and a process for flagging underperformance, you have no leverage after deployment.
  
• Name a clinical owner for every AI output. Not a department. A specific person. An alert that belongs to everyone gets acted on by no one. Assign which role receives which output and what happens within what timeframe.
  
• Set a 90-day review point in the contract. Most AI tools look promising in demos and controlled pilots. Real-world performance in your patient population, with your staffing, inside your EHR, is a different question. Build in a structured review before the tool becomes permanent infrastructure.
  
• Plan for the bias audit. As AI tools become more embedded in clinical workflows, US hospitals will face increasing regulatory and ethical pressure to demonstrate that their AI tools perform equitably. Getting ahead of this now, by reviewing subgroup performance data and building audit processes, is less expensive than addressing it after a patient harm event.

Final Thoughts

The cardiologist who caught the silent AF. The algorithm that flagged the patient as high-risk for readmission three days before discharge. The AI scribe that gave back 45 minutes at the end of a 12-hour shift. These are not hypothetical futures. They are happening in US hospitals today.

But so is the AF algorithm that was never validated on a diverse population. The alert fatigue system that fires 200 low-priority notifications a day. The AI dashboard that no one opens.

The technology is real. The results are uneven. The difference is almost always in how the tool is selected, deployed, and connected to care.