Category: Artificial Intelligence

Google’s AMIE Shows Promise for Cardiology Care

Posted on by Viktor Zarev

AI scribes and LLMs’ potential impact on cardiology just became a lot clearer after a Stanford RCT published in Nature Medicine demonstrated that Google’s AMIE system helped general cardiologists better diagnose and plan treatment for rare heart conditions.

  • The WHO predicts an 18M healthcare provider deficit by 2030, with the shortage hitting hardest in subspecialties like those across cardiology.
  • Large language models might help alleviate this shortage, but rigorous LLM RCTs are rare across medical specialties with zero RCTs evaluating LLMs in cardiology until now.
  • Among the healthcare LLMs is Google’s AMIE, which focuses on gathering conversational information and interpreting medical scans like lab tests and ECGs.

Researchers enrolled nine general cardiologists to evaluate 107 patients with suspected genetic cardiovascular disease. Each case was then reviewed by two cardiologists (one with access to AMIE and one without) and three subspecialists later confirmed the diagnoses.

  • AMIE-assisted assessments were more often supported by the subspecialists (46.7% vs 32.7%).
  • As were the resulting management plans (45.8% vs 29.9%), 
  • And diagnostic recommendations (43.9% vs 30.8%).

When it came to catching mistakes…

  • Clinical errors occurred in 13% of AI-assisted versus 24% of unassisted assessments.
  • Meanwhile general cardiologists reported AMIE improved their assessments in 57% of cases, increased confidence in 52%, and saved time in 50%.

But AMIE also made some mistakes of its own, with hallucinations occurring in 6.5% of cases.

  • The hallucinations were mild (such as assuming patient gender or fabricating minor imaging findings), and they often self-correcting when challenged by cardiologists.

What is perhaps most remarkable about the study is that AMIE adapted to this level of subspecialist cardiology with considerable data efficiency, requiring feedback from experts on just nine cases.

Taking it all into consideration in their discussion, the study’s authors suggest that AMIE is still better at improving treatment planning than initial diagnostic reasoning, with that part still best left to expert physicians.

The Takeaway

This study is notable in the cardiology community as the first RCT to truly evaluate how an LLM can support general cardiologists with handling complex CV conditions that usually belong to subspecialists. The results suggest it’s a strong starting point that could help improve treatment planning and management, while the verdict is still out on diagnostics.

How AI is Expanding ECG and Echo’s Applications

Posted on by Viktor Zarev

Building stronger evidence for AI’s place in cardiology, this week brought several examples of how AI can improve ECG use cases and automate Echo for new disease detection methods.

Starting with AI improving ECG for coronary revascularization, a deep learning ECG model outperformed both clinician-led ECG and troponin T testing in predicting which ED patients needed coronary revascularization. 

  • The model was trained on 145k U.S. cases and achieved an AUROC of 0.91 vs. 0.65 for clinician reads and 0.71 for TnT. 
  • External validation in Europe (18k patients) showed strong results too, with an AUROC of 0.81 for revascularization and 0.85 for type 1 MI.

AI could also expand ECG into HF risk prediction as part of a recent JAMA study which found that a positive AI-ECG screening result predicted a potentially seven-fold higher HF risk.

  • AI-ECG’s discrimination for new-onset HF was 0.723 in the U.S., 0.736 in the U.K., and 0.828 in Brazil. 
  • When compared to the standard PCP-HF and PREVENT equations for classifying HF, AI-ECG showed net reclassification improvements of up to 47.2% and 47.5%.

A third study gave another example of how AI can automate echocardiography by using deep-learning to accurately detect clinically significant tricuspid regurgitation from full transthoracic echoes.

  • The AI model detected moderate or severe TR with an AUC of 0.95 and severe TR with an AUC of 0.98.
  • Researchers now hope this project can serve as the foundation for future AI–assisted Echo workflows in TR patients thanks to the model’s strong accuracy.

The Takeaway

While these studies might not sway all the AI skeptics in healthcare, they’re part of a growing trend that shows how AI can expand the functionality and efficiency of well-established modalities like ECG and Echo.

AI-ECG’s 2025 Kickoff

Posted on by Viktor Zarev

The buzz around artificial intelligence’s impact on cardiology keeps growing louder, and that’s proving to be particularly true in the AI-ECG segment, with 2025 already off to a strong start from ECG startups and researchers alike.

  • AI and ECG pair well due to the strong pattern recognition that machine learning algorithms can achieve.
  • This AI-ECG pairing is bringing the modality into a wider range of diseases and allowing integration with new device form factors.

One pertinent example is AccurKardia’s recent FDA breakthrough designation for its ECG-based, AI-powered AK+ Guard hyperkalemia detection software. 

  • AK+ Guard uses Lead I ECG data to alert patients and clinicians of moderate to severe episodes of excess blood potassium (hyperkalemia) that can lead to sudden cardiac arrest.
  • Accurkardia designed AK+ Guard for consumer and clinical wearables to support remote hyperkalemia monitoring for patients with renal disease, CKD, and other risk factors.
  • For reference, 37M people in the U.S. suffer from CKD, and hyperkalemia is associated with a 16.6% higher mortality rate in those patients.

Tackling another CV complication, researchers developed an AI-ECG risk estimator for hypertension (AIRE-HTN) that could become a useful tool for predicting future CV events. 

  • AIRE-HTN was trained on ~190k patients at Beth Israel Deaconess and validated on 65k patients from UK Biobank.
  • Ultimately, the study found that patient AIRE scores accurately predicted CV death (HR: 2.24), HF risk (HR: 2.60), MI (HR: 3.13), ischemic stroke (HR: 1.23), and CKD (HR: 1.89) compared to traditional risk factors.

The commercial side of the AI-ECG arena is similarly heating up, including a new partnership between AliveCor and ECG AI developer Anumana.

  • The partnership will focus on AliveCor’s Kardia ECG devices and Anumana’s ECG-AI algorithms, and apparently includes both AI development and integration.
  • Their first target will be Anumana’s ECG-AI algorithm, which detects low ejection fraction using 12-lead ECG data, and recently landed CMS reimbursement.

The Takeaway

Though there’s no shortage of AI skeptics in healthcare, it’s becoming clear that AI is bringing ECG into new disease areas and form factors. For a modality as established as ECG, that’s a big deal, and it could have a major impact on both disease detection and patient access.

Cardiology AI Clearances Growing, Diversifying

Posted on by Jake Fishman

The FDA updated its healthcare AI database last week, increasing its list of AI-enabled clearances to a whopping 950 medical devices, while highlighting some interesting trends in cardiovascular AI.

Overall healthcare AI clearances seem to be stable, with the first half of 2024 bringing 107 total clearances, which is just behind pace of 2023’s full year total (220).

Cardiovascular AI maintained a distant second largest share of FDA-cleared AI products in H1 2024, with 9% of total clearances (10 devices), well below radiology’s 73% share (78 devices).

However, cardiovascular AI actually made up a larger 18% share of total H1 2024 clearances (19 devices) if you also count cardiovascular imaging AI products that the FDA technically categorized within its “Radiology” segment (e.g. FFRCT, coronary plaque, echo AI, etc)…We’re using this broader definition of cardio AI through the rest of this story.

  • Cardiovascular AI’s total share of H1 2024 AI clearances was the highest since 2020 (both 18%), after landing between 14% and 16% during the last three years.
  • Cardiovascular AI’s 19 clearances in H1 2024 also puts the segment on pace to eclipse any previous year (previously 30, 24, 21, 20, 22, 15).

The FDA database also reveals extremely wide brand diversity in the cardiovascular AI segment, with 86 companies, and none with more than a 6% share of total clearances.

  • That trend stayed alive in H1 2024, which saw another influx of new cardio AI developers making their debut on the FDA list (AgileMD, DASI Simulations, inHEART, InVision Medical, Innolitics, Kestra Medical… and even the US Army).

Cardiovascular AI applications continued to largely analyze imaging and ECG data, although imaging+ECG’s share of cardio AI clearances fell to 74% in H1 2024 (down from 83% in 2023 and 88% in 2022) due to a recent increase in EHR and sensor-based cardio AI products.

  • We’re also seeing cardiovascular AI continue to expand from diagnostics/detection to more procedural use cases, including new solutions for TAVR and ablation planning that were cleared in H1 2024.

The Takeaway

Although actual cardiovascular AI use in the clinic is in its early stages, the large and growing list of FDA-cleared cardio AI products is a reminder of the innovations taking place in this arena. That innovation appears to be leading to more cardiovascular AI products, with more diverse use cases, and should eventually lead to larger increases in clinical adoption.

New Echo AI Model Could Streamline Mitral Regurgitation Diagnosis

Posted on by Jake Fishman

Columbia and Cornell researchers developed an echo AI model that could improve the difficult task of mitral regurgitation diagnosis, and might even represent an AI-driven step towards enhancing all valvular regurgitation assessments.

  • Transthoracic echocardiography is the go-to modality for mitral regurgitation diagnosis, but these exams are difficult to perform and prone to high variability.

The researchers set out to solve this problem using AI, and indeed showed that AI could perform MR exams at the same level as academic echo labs.

Using 52,702 Columbia-sourced TTEs for model training and validation, the researchers developed their end-to-end AI system to intake complete TTE studies, identify color MR Doppler videos, then accurately determine MR severity on a 4-step scale (none/trace, mild, moderate, and severe). 

They then tested their AI against 8,987 TTEs from an internal Columbia dataset and 8,208 TTEs from an external Cornell dataset, finding that the model achieved high…

  • Agreement with cardiologist interpretations – 82% & 79% accuracy (k = 0.84 & 0.80)
  • Performance for detecting moderate/severe MR – 0.98 & 0.98 AUROCs

Most misclassification disagreements between the AI and cardiologists involved exams with none/trace or mild MR, while the AI maintained robust performance across different MR types (slightly lower with eccentric MR cases).

  • When the AI and cardiologist disagreed, an adjudication panel sided with the AI and cardiologists about a quarter of the time each, while half the time they thought the answer was in the middle.
  • The researchers also found that AI trained to use multiple TTE views outperformed models using the apical 4-chamber view (82% vs. 80%).

Next up, the researchers plan to continue their echo AI efforts, including expanding to other valvular and cardiac assessments, and bringing their research into clinical settings.

The Takeaway

Echocardiography has long been an AI hotspot due to its importance, prevalence, and its challenges with efficiency and variability, but most clinical-level echo AI solutions have focused on heart failure and aortic stenosis so far. Although Columbia and Cornell’s new AI model is still in its early stages, this study makes a solid case for how much more echo AI can help, including across a full range of valvular regurgitation assessments.

Philips Expands & Integrates Echo AI Platform

Posted on by Jake Fishman

Philips’ ultrasound AI strategy took another big step this week, with the launch of its next-generation echo AI platform, which will come integrated with the company’s cardiovascular ultrasound systems and bring a range of new echo-automating capabilities. 

Philips’ new AI-enabled cardiovascular ultrasound platform includes a combination of established and first-to-market AI applications that come fully integrated with its EPIQ CVx and Affiniti CVx echo scanners, with a focus on speeding up echo teams’ cardiac ultrasound analysis.

The announcement specifically highlighted Philips’ first-of-their-kind…

  • Segmented Wall Motion app – an automated tool for segmental wall motion scoring and identifying disorders such as coronary artery disease and cardio-oncology issues.
  • 3D Mitral Regurgitation Volume Quantification app – a tool for the automated analysis of MR flow dynamics, supporting valve treatment decision-making (not yet FDA cleared).

The platform also includes solutions for automated LVEF measurements, LV strain analysis, and the automated selection of the most appropriate LV images.

Philips has been developing echo AI solutions for some time through its ultrasound and Tomtec teams, and boosted its echo AI portfolio and development capabilities through its acquisition of DIA Image Analysis almost exactly one year ago.

However, this launch is perhaps most notable for bringing the “deep” integration of AI tools into Philips’ cardiovascular ultrasound systems, making AI a core part of its scanners’ feature set and value proposition, rather than an add-on.

  • That integration could be particularly notable from a usability perspective, and could also help drive AI adoption given the continued barriers experienced when selling / buying AI as standalone solutions. 
  • It might also be a sign of an echo AI integration trend, noting that GE HealthCare has been steadily expanding its Caption AI echo guidance solution to more of its echo scanners since acquiring Caption Health last year.

The Takeaway

With all the imaging AI acquisitions that have taken place over the last few years, Philips’ echo AI integration is a reminder why the only acquisitions made by the big OEMs targeted cardiovascular ultrasound AI solutions (no other indications or modalities). 

AI tools like these make cardiovascular ultrasound systems faster and easier to use, and scanner integrations like these might have a multiplying effect on echo AI’s adoption and overall impact.

AI Clearances Surge, Cardio AI Share Declines

Posted on by Jake Fishman

The FDA published its latest Healthcare AI Database last week, featuring a massive 882 AI-enabled medical device clearances, and highlighting some interesting trends in cardiovascular AI.

Overall healthcare AI clearances are gaining momentum, with 2023 bringing a 42% jump in clearances, more than doubling 2022 and 2021’s annual growth (+20% & +16%). 

  • However, this growth is in part due to the fact that products must get re-cleared as their algorithms change, and the growth of unique AI products is far more modest.

Cardiovascular AI maintains a (distant) second largest share of FDA-cleared AI products, with 10% of total clearances (90), well below radiology’s 76% share (671).

Cardiovascular AI actually makes up a larger 17.4% share of total clearances (154) if you also count cardiovascular imaging AI products that the FDA technically categorized within its “Radiology” segment (e.g. FFRCT, coronary plaque, etc).

  • Even with this broader definition, cardiovascular AI’s total share of AI clearances is declining, falling from roughly 25% of clearances in 2018-2019, to 16.5% in 2020-2022, and 13.5% since the start of 2023.
  • Cardiovascular AI’s falling share is partially due to a surge in AI products from new specialties (orthopedics, pathology, urology, ENT), but it’s mainly because non-cardiac radiology AI applications scored a whopping 544 clearances since 2020.

Cardiovascular AI applications also appear to be getting more diverse. Between 2020 and 2022, an overwhelming 86% of all cardiovascular AI clearances were for products that either analyzed imaging or ECG, but imaging and ECG AI’s share of cardiovascular AI clearances fell to 66% in 2023-2024. 

  • We’re also seeing cardiovascular AI expand from diagnostics/detection to more procedural use cases, including a growing number of EP ablation mapping products and even AI-enhanced cardiac implants.

The Takeaway

Although the massive growth in healthcare AI clearances far outpaces the actual use of AI in healthcare, these trends are still remarkable and a testament to the huge potential healthcare leaders see in artificial intelligence.

This is of course also true for cardiovascular AI, which might be seeing its share of overall clearances declining, but is still home to some of the most-funded and most-used AI startups, as well as some of the most innovative and clinically relevant AI use cases.

Cedars-Sinai’s Massive Echo + Reporting Foundation Model

Posted on by Jake Fishman

The team at Cedars-Sinai’s Smidt Heart Institute has become among the most prolific health system-based cardiovascular AI developers, and they added to their AI resume after unveiling the massive new EchoCLIP foundation model, which combines echo images and reporting text to perform a wide range of interpretations — without specific training.

  • Foundation models are a type of generative AI, using a vast amount of unlabeled data to perform a wider range of clinical tasks than we’ve seen with most current healthcare AI tools.

To develop this beast of an echo AI model, the team assembled a dataset of 1,032,975 cardiac ultrasound videos and corresponding text-based expert interpretations.

Even without task-specific training or fine-tuning, EchoCLIP performed well across a wide range of measurement and detection tasks when tested against external data, including:

  • Accessing cardiac function by predicting LVEF (7.1% mean absolute error)
  • Estimating pulmonary artery pressure (10.8 mm Hg mean absolute error)
  • Identifying intracardiac devices, like mitral valve repair, TAVR, and pacemaker/defibrillator leads (AUCs = 0.97, 0.92, 0.84)
  • Detecting changes from a healthy cardiac chamber size, like severe dilation of the right ventricle, right atrium, left ventricle, and left atrium (AUCs = 0.92, 0.97, 0.92, 0.91)
  • Assessing tamponade and severe left ventricular hypertrophy (AUCs = 0.96 & 0.82)

Perhaps more impressively, the team’s related EchoCLIP-R system was able to accurately identify specific patients using only their exams and retrieve past exams (AUC = 0.86), while highlighting clinically important changes that occurred between their echos.

Altogether, these results suggest that with a large enough dataset of echo images and expert text interpretations we can train foundational models that can support an extremely wide range of echo assessment tasks. 

The Takeaway

The last few years have brought an impressive flow of echo AI models, and Cedars-Sinai’s new EchoCLIP model certainly could prove to be among the most significant given its size, breadth of capabilities, and its role as the first of potentially many advanced echo image+text foundation models.

Cedars-Sinai Detects AFib with Echo AI

Posted on by Jake Fishman

A Cedars-Sinai-led team developed an echocardiography AI model that accurately detects patients with atrial fibrillation – including asymptomatic patients – potentially revealing a new “opportunistic screening” method to drive earlier AFib detection and treatment.

  • Transthoracic echo (TTE) is the most common cardiovascular imaging exam, and is increasingly being combined with AI to support image acquisition, interpretation, and reporting.
  • AFib detection has historically relied on ECG exams, and although ECG AI has shown promise, many patients who undergo ECGs were already either suspected to have AFib or already experienced a major cardiac event.

The researchers trained their deep learning algorithm using 111k TTE videos, including 39k exams from patients who were in AFib and 72k exams from patients with normal sinus rhythm at the time of their echos… although 6,654 of those “normal” exams were from patients who actually had paroxysmal AFib.

  • When tested against an internal Cedars-Sinai TTE dataset, the model identified patients who were in AFib with “high accuracy” (AUC: 0.96) and patients with normal sinus rhythm but paroxysmal AFib “moderately well” (AUC: 0.74).
  • When tested against 10k TTEs from an external Stanford dataset, the model achieved a decent 0.69 AUC identifying patients with a history of AFib.

That performance was consistent across genders, older patients, and patients with higher AFib risks, and perhaps more notably, it outperformed AFib detection based on clinical risk factors, TTE measurements, LA size, and CHA2DS2VASc Scores (AUCs: 0.64, 0.64, 0.62, 0.61).

They then combined the TTE AI model with an ECG-based deep learning model, finding that the combined TTE/ECG system detected patients with AFib better than the ECG model on its own (AUC 0.81 vs. 0.79), potentially because structural information from TTEs might complement ECG-based analysis. 

The Takeaway

There are about 6 million people in the United States living with AFib, and nearly a quarter of them are currently undiagnosed. Although Cedars-Sinai’s echo AI AFib model is still in its early stages, this study makes a solid case for how AI could be used to proactively analyze the 7 million echo exams that are performed in the US each year to help identify more undiagnosed AFib patients while their treatment would be most effective.

Chest X-Rays’ Cardiovascular Screening Potential

Posted on by Jake Fishman

There are over 70 million chest X-rays performed in the U.S. every year, and a pair of new studies highlighted how AI could be used to “opportunistically screen” these exams for undetected cardiovascular disease.

A Mass General Brigham team writing in the Annals of Internal Medicine detailed how their deep learning system was able to identify patients with greater 10-year risks of experiencing major cardiovascular events by analyzing their chest X-rays.

The researchers developed their ‘CXR CVD-Risk’ model using chest X-rays sourced from a cancer screening trial, and then externally validated the AI model against 11,000 patients’ CXRs.

  • Among the patients with unknown ASCVD risks, those who received a 7.5% or higher CXR CVD-Risk estimate had a 73% greater adjusted 10-year risk of MACE than patients with estimates below 7.5%.
  • Among the patients with known ASCVD risks, the AI model added to the MACE-prediction accuracy of their traditional ASCVD risk scores (adjusted HR = 1.88).

A Columbia University team writing in the European Heart Journal showed that CXR deep learning algorithms can also be used to identify patients with left ventricular structural abnormalities that might be signs of heart failure.

The researchers developed their algorithm using 71,589 unique CXRs from 24,689 patients, along with the patients’ echocardiogram labels for two structural abnormalities: left ventricular hypertrophy (SLVH) and dilated left ventricle (DLV). 

  • When tested against 3,667 CXRs from an internal Columbia dataset, the model detected patients with SLVH,  DLV, or either of the abnormalities with relatively high AUCs (0.79, 0.80, 0.80).
  • When validated against 8,003 CXRs from an external Stanford dataset, the model detected patients with SLVH, DLV, or either of the abnormalities with lower but still decent AUCs (0.67, 0.78, 0.76).
  • The model also outperformed 15 board-certified radiologists’ visual assessments to detect patients with either SLVH or DLV, with 71% sensitivity (vs. 66%) at a fixed specificity of 73%.

The Takeaway

We’ve already seen significant progress in using chest CT scans to opportunistically screen for patients with undetected coronary artery calcium. But considering the massive volume of chest X-rays performed each year, solutions like these could have a major impact – especially as algorithm accuracy and post-detection workflows continue to evolve.

Get twice-weekly insights on the biggest stories shaping cardiology.