AI + EKG: Old Wires, New Tricks

he humble electrocardiogram has been around for more than a century—Einthoven’s string galvanometer made its debut in 1903—and despite generations of new monitors and algorithms, the core idea hasn’t budged. It’s still just a summation of electrical vectors flickering across the myocardium, projected onto a piece of paper (or a digital screen). And yet, this century-old signal may still be holding secrets we’ve never seen.

Recent work suggests the ECG might whisper early clues about heart disease long before structural or symptomatic changes appear—findings already noted in conditions like cardiac amyloidosis and pre-HF states. With deep learning, models can now comb through millions of waveforms and spot subtle signatures invisible to human readers.

But the latest Circulation study from Yale’s CARDS Lab (Croon et al., 2025) offers a reality check. The researchers examined more than 234,000 patients across tertiary and community hospitals, outpatient clinics, and the UK Biobank, testing five disease-specific AI-ECG models (for LV systolic dysfunction, aortic stenosis, mitral regurgitation, LV hypertrophy, and a composite structural-heart-disease model). Each model was compared against diagnoses and future cardiovascular events over five years.

The punch line: the models didn’t stay in their lanes. Their “phenotypic fingerprints” overlapped almost completely (correlations 0.67–0.96) and lit up across a broad spectrum of heart conditions—not just the one they were trained to detect. Even so, these models predicted future HF and other CV events impressively well. The message? AI-ECG may function less like a yes/no test for one pathology and more like a generalized cardiovascular biomarker—a signal to look deeper rather than to label.

That’s where imaging may pick up the baton. In a Nature study introducing EchoNext showed that AI applied to routine 12-lead ECGs could triage who actually needs an echocardiogram, reaching AUROCs ≈ 0.8 across multiple systems and outperforming cardiologists.

Put together, these papers mark a turning point. The ECG—once the simplest test in cardiology—may become the front door to a multimodal AI ecosystem, flagging risk, guiding imaging, and contextualizing findings across platforms. Still, the road to precise, disease-specific AI remains long. For now, think of AI-ECG not as a diagnosis, but as a smart red flag that tells you when it’s time to listen closer.

Share DrAiMD’s Substack

AI + Cardiac Imaging: From Pixels to Prognosis

For decades, interventional cardiology has chased a tantalizing goal: if we can see a plaque, we can stop a heart attack. When coronary stents hit the scene in the 1990s, many hoped the era of acute MI might soon end. But landmark trials like COURAGE, FAME, and ISCHEMIA reminded us that even with high-resolution angiograms, we’re often fixing the wrong lesions. Angiography shows the narrowest spot, not necessarily the next culprit. Vulnerable plaques—thin-capped, lipid-rich, inflamed—often lurk quietly before they rupture.

That’s why the newest work in optical coherence tomography (OCT) is so compelling. OCT uses near-infrared light to deliver microscopic, histology-like images of the coronary wall in vivo. Until recently, analyzing those images took hours of expert reading. Enter AI.

In Circulation: Cardiovascular Imaging, researchers introduced AutoOCT, a deep-learning system that cleans artifacts and quantifies plaque features—fibrous-cap thickness, lipid arc, calcium burden—across full-length pullbacks in minutes. It matched expert and core-lab performance, detected statin-related plaque stabilization, and highlighted features tied to future major adverse events.

The PECTUS-AI study in the European Heart Journal took the next step: using OCT-AID to automatically flag thin-cap fibroatheromas in post-MI patients. When the AI analyzed entire arterial segments—not just culprit lesions—it nearly doubled prognostic accuracy, with a 97 % NPV for near-term events when no TCFA was found. That kind of risk stratification could one day guide who truly benefits from early intervention rather than empiric stenting.

Outside the cath lab, imaging AI is expanding fast. On stress cardiac MRI, a fully automated model now measures global strain and independently predicts 5-year MACE, outperforming traditional EF and ischemia metrics.

And in echocardiography, EchoNet-Measurements—built on nearly 900,000 prior echoes—automates 18 standard B-mode and Doppler parameters with sonographer-level accuracy, releasing open-source code and overlays clinicians can verify.

Together, these advances hint at a new imaging era. AI won’t replace angiography or the human eye, but it may finally show which plaques—and which patients—truly sit on the edge of catastrophe. At the same time, automation in echo and CMR could make high-precision imaging cheaper, faster, and more accessible, bringing sophisticated risk assessment to hospitals and clinics far beyond academic centers.

The tools are learning fast; the challenge now is making sure we do, too.

AI + Regulation & Market: Oversight in the Age of Learning Machines

We at AIMD are unabashedly optimistic about AI in healthcare. But optimism shouldn’t mean skipping due diligence. The conversation around artificial intelligence in medicine isn’t just about algorithms and AUCs anymore—it’s about regulation, safety, and how innovation actually reaches the bedside.

A new JACC: Advances review took a bird’s-eye view of the market and found more than 200 FDA-reviewed cardiovascular AI/ML products cleared through 2024. The trend line is almost vertical—85% approved since 2018—but the fine print matters. Roughly 96% of those clearances came through the 510(k) pathway, a regulatory route designed to show that a new device is substantially equivalent to one already on the market.

The 510(k) process is faster and less burdensome than a de novo or full premarket approval—it’s essentially a noninferiority standard. That’s ideal for incremental improvements, but less so for tools claiming to change clinical decision-making. Only a small fraction of these AI devices had prospective or trial-level validation, and fewer still reported detailed demographic data. It’s a gentle reminder that while the AI market is booming, rigorous, real-world evidence hasn’t quite caught up.

Meanwhile, the FDA’s new Predetermined Change Control Plan (PCCP) guidance marks a major step forward for adaptive AI. It allows manufacturers to predefine how models can update themselves—what data they’ll use, how retraining will be verified, and what safety checks are in place—without filing a new submission each time. In theory, it keeps regulation aligned with how AI actually evolves.

In practice, though, it raises familiar questions: What if the data feeding those updates are biased or unrepresentative? What if performance drifts faster than postmarket monitoring can detect? As with most things AI, flexibility comes with risk.

Still, this is an essential evolution. If we want AI tools that truly learn over time, regulators have to meet them halfway. For clinicians, this moment signals a new literacy to master—understanding not just the biology behind our diagnostics, but the regulatory DNA of the algorithms we deploy. Knowing how a model was trained, tested, and approved may soon be as fundamental as knowing the evidence behind a drug or the inclusion criteria of a trial.

Because in the age of learning machines, oversight is part of clinical judgment.

AI-ECG isn’t picky about what it finds

Turns out heart-focused AI models aren’t so narrow—they light up across the cardiovascular spectrum, acting more like general risk detectors than one-disease specialists.

A new Circulation study led by Rohan Khera’s CARDS Lab at Yale suggests that artificial intelligence–enhanced ECG models may be less like sharp scalpels and more like wide nets. These algorithms, often trained to spot a specific condition—say, left ventricular systolic dysfunction (LVSD)—don’t just detect their target. They also flag a broad spectrum of cardiovascular problems with nearly equal strength.

PICO snapshot:

• Population: 233,000+ patients across Yale health systems and the UK Biobank.

• Intervention: Six AI-ECG models trained to detect LVSD, aortic stenosis, mitral regurgitation, LV hypertrophy, or composite structural heart disease.

• Comparison: Traditional phenotype associations and negative-control AI models (e.g., biological sex, non-cardiac conditions).

• Outcome: On- and off-target disease associations; new-onset CVD risk prediction.

Key findings:

• Phenotypic signatures were highly overlapping (correlation r = 0.67–0.96).

• Off-target associations (like ischemic cardiomyopathy) often exceeded on-target strength.

• Across cohorts, AI-ECG predictions consistently forecasted future cardiovascular disease, especially heart failure.

Takeaway: These tools may work best as broad biomarkers of cardiovascular vulnerability, not as a binary diagnostic modality.

AI-Powered Stethoscope (TRICORDER Study)

A 15-second diagnosis sounds great on paper—but high false positives and clinician dropout show that adoption needs more than accuracy.

Researchers from Imperial College London and Imperial College Healthcare NHS Trust introduced a novel AI-enabled stethoscope—developed in partnership with Eko Health—that simultaneously captures ECG and heart sounds and processes them via AI to detect heart failure, atrial fibrillation (AF), and valvular disease in just 15 seconds. This tool was trialed in about 12,725 symptomatic patients across 96 GP surgeries and compared to controls from 109 surgeries. Findings included:

• 2.33× likelihood of heart failure diagnosis,

• 3.45× likelihood of atrial fibrillation diagnosis,

• 1.92× likelihood of valvular heart disease diagnosis within 12 months, compared to non-AI-assessed patients .

The trial revealed limitations too: around 2/3 of AI-flagged heart failure cases were false positives, and 70% of practices ceased using the device within a year, highlighting challenges in integration .

Other Scholarly Work

Zhou et al. published an AI stethoscope study in JACC: Advances (2025), demonstrating diagnostic performance for left-sided valvular heart disease using the Smartho-D2 electronic stethoscope. With 514 patients, they trained on 304 and tested on 210: AU-ROC for composite VHD was 0.85, with sensitivity ~70% and specificity ~74%, and diagnosis accomplished in ~91 seconds—faster than standard echocardiography .

Bottom Line

These results signify that AI-driven auscultation may soon augment the traditional physical exam, especially in resource-limited or time-constrained settings. The balance of sensitivity vs. false positives and implementation barriers (and ultimately clinician buy-in) remain barriers.

The critical question: Can AI tools re-dignify the bedside exam by fusing tradition and precision? Or will they become another abandoned gadget, outpaced by imaging and lab advances?

EU prescribes a dose of reality for AI in Healthcare

A sweeping EU report finds that messy data, regulation, and trust gaps are slowing AI’s clinical impact—though Europe may be uniquely positioned to set the gold standard.

The European Commission dropped a hefty report on AI in healthcare, and the takeaway is clear: plenty of promise, not much practice. Despite a growing market of AI tools, real-world adoption is sluggish thanks to messy data systems, complex regulation, funding gaps, and clinicians wary of black-box models.

The study points to global lessons—like Israel’s national data frameworks and Japan’s clinician training—that Europe could emulate. Proposed next steps include EU-wide data standards, centres of excellence, new financing models, and even an AI “catalogue” to guide hospitals.

Bottom line: The EU says it can strike the balance between innovation and rights. But unless integration gets smoother, AI risks being more hype than healing.

Audio version powered by NotebookLLM

Subscribe now

Share DrAiMD’s Substack

Julius AI: Revolutionizing Data Analysis for All Skill Levels

Julius AI is changing the game in data analysis by making advanced tools accessible to everyone, regardless of expertise. Whether you’re handling business analytics, healthcare data, or marketing stats, Julius AI turns raw data into actionable insights through intuitive charts, graphs, and detailed reports.

Ready to dive in? Sign up using this link and start transforming your data today!

How to Get Started:

• Upload your data file or connect to Google Sheets.

• Start a conversation with Julius to analyze and visualize your data using simple, natural commands.

• Generate insights instantly, whether you need complex sorting or easy-to-understand bar charts.

Pricing Plans:

• Basic: $20/month, perfect for light users.

• Essential: $45/month, the most popular plan with unlimited messages.

• Team: $70/month per member, built for organizations.

AI in Academic Publishing: Policies from Elsevier, Springer Nature, and Cureus

As AI tools like ChatGPT gain traction in academic research, top journals are setting boundaries to ensure responsible use. Let's dive into the AI-related policies of three major publishers—Elsevier, Springer Nature, and Cureus—focusing on AI authorship, content creation, and peer review.

AI authorship is a no-go. All three publishers prohibit listing AI as an author since it lacks the capacity for responsibility. Elsevier, Springer Nature, and Cureus allow AI to assist with grammar and readability, but content generation must be human-driven.

Language vs. research content: AI can polish writing, but it’s strictly banned from fabricating research conclusions. Cureus, for instance, is clear that while AI can edit, it shouldn’t alter intellectual content.

AI in peer review? Not yet. These publishers agree that AI isn’t ready to replace human expertise in evaluating research due to bias and confidentiality risks.

The bottom line is that AI is a helpful tool for researchers, but human accountability remains non-negotiable.

AI & Health Info: Trust Issues?

The latest KFF Health Misinformation Tracking Poll sheds light on how Americans interact with AI for health information. While two-thirds of adults have used or interacted with AI, only one-third do so regularly. Surprisingly, a majority (56%) say they lack confidence in their ability to distinguish fact from fiction when using AI chatbots for health advice.

AI + Health = Trust Issues
A mere 29% of adults trust AI bots like ChatGPT to provide reliable health info. This trust is higher for tasks like tech advice (48%) or home maintenance tips (54%), but skepticism dominates the health sector. Even among AI users, just 36% trust AI for accurate health advice.

Is AI helping or hurting health information seekers? The jury's still out—55% of adults say they aren’t sure of its impact, while 21% think it helps, and 23% believe it hurts efforts to find accurate information.