How Machine Learning is Applied in Cardiology

Yann LeCun's vision: machines should learn like humans — by building internal world models, not reconstructing every pixel. We just validated this idea at the largest scale ever attempted in cardiac ultrasound. Introducing EchoJEPA — the first world model for medical video.🔥 🫀 18M echocardiograms 👥 300K patients 🧠 Learns heart dynamics — not imaging noise The problem: Ultrasound is messy. Speckle, shadows, attenuation. Most pretraining objectives end up modeling the scanner, not the heart. The idea: Stop reconstructing pixels. Predict latent structure instead. EchoJEPA discards what’s unpredictable and locks onto what matters clinically: ➡️ chamber geometry ➡️ wall motion ➡️ valve dynamics The results (frozen encoder, no fine-tuning): • 20% ↓ error in LVEF • 17% ↓ error in RVSP • 79% accuracy with 1% labels (vs 42% for baselines w/ 100%) • 2% degradation under acoustic artifacts (vs 17%) • Zero-shot pediatric transfer beats all fine-tuned models Why this works: When we project embeddings: ❌ prior methods → diffuse, entangled clusters ✅ EchoJEPA → clean anatomical organization Structure separated from acquisition noise. 📄 Paper: https://lnkd.in/gPxhQpCR 💻 Code: https://lnkd.in/gQ-i6yMx Huge credit to Alif Munim, who pushed JEPA thinking into medical video and led this effort 💥 Guidance from AI at Meta (Quentin Garrido, Koustuv Sinha) Co-authors: Adib Fallahpour Teodora Szasz Ahmadreza Attarpour, PhD etc! Teams: University Health Network Amazon Web Services (AWS) University of Toronto UChicago Medicine University of California, San Francisco Philips This is representation learning for physiology, not pixels.

𝗔𝗜 𝗖𝗮𝗻 𝗡𝗼𝘄 𝗣𝗿𝗲𝗱𝗶𝗰𝘁 𝗛𝗲𝗮𝗿𝘁 𝗙𝗮𝗶𝗹𝘂𝗿𝗲 𝗙𝗶𝘃𝗲 𝗬𝗲𝗮𝗿𝘀 𝗕𝗲𝗳𝗼𝗿𝗲 𝗜𝘁 𝗗𝗲𝘃𝗲𝗹𝗼𝗽𝘀 A fascinating paper published this week in the The American Journal of Cardiology. The important question it asked was the following. Can we predict heart failure before the heart has already begun to fail? The answer, it now appears, is yes. A team led by Prof Charalambos Antoniades MD PhD FRCP FMedSci at the University of Oxford has developed an AI tool that analyses the fat surrounding the heart from routine cardiac CT scans, predicting a patient's risk of developing heart failure up to five years before any clinical signs appear. Epicardial adipose tissue (EAT) is a metabolically active visceral fat depot that is both a sensor and a modulator of myocardial biology and changes its composition in response to paracrine signals from the myocardium. The team hypothesised that radiomic characterization of EAT from routine coronary computed tomographic angiography (CCTA) can noninvasively capture this adverse remodeling and enable early heart failure (HF) risk stratification. The study involved over 72,000 patients across nine NHS centres, followed for up to a decade. The fat around the heart, it turns out, acts as a potential biological sensor. Patients in the highest risk group were twenty times more likely to develop heart failure than those in the lowest. The tool predicted five-year risk with 86% accuracy, outperforming models built on traditional risk factors alone. What is striking is the conceptual shift this represents. We have spent decades in cardiovascular medicine treating disease that has already declared itself, responding to symptoms, managing complications, optimising a heart already under strain. We have been using risk stratification of cardiac disease using various methods like calcium scores. The team are now seeking NHS regulatory approval and adapting the tool for any CT scan of the chest, not just cardiac ones. Every scan, for any reason, could soon carry an embedded layer of cardiac risk intelligence. As the NHS shifts into prevention as part of the long term plan these tools become more important.

Case Tuesday: Cardiac CT A patient presents with chest pain. The question is urgent: is this a heart attack waiting to happen, or something else? A CT coronary angiogram is performed. For the radiologist, this means carefully assessing coronary arteries, looking for stenosis, calcifications, and subtle plaques. The challenge: Coronary CTs generate hundreds of slices, often complex to interpret. Subtle plaques can be easily overlooked. Quantifying calcium scores and stenosis consistently takes significant time. This is where #AI is showing real promise: Automated calcium scoring to assess cardiovascular risk Plaque detection and quantification to support precise diagnosis Tools that standardize reporting and improve communication with cardiologists The radiologist’s expertise is essential in interpretation and clinical context but AI ensures that the assessment is faster, more reproducible, and more actionable. The impact: Earlier detection of coronary artery disease. Better risk stratification for patients with chest pain. Closer collaboration between radiology and cardiology teams As Chief Medical Officer at GE HealthCare, I see cardiac CT as a shining example of how AI doesn’t just enhance workflows it helps us move toward preventive, precision medicine that saves lives before catastrophe strikes. Do you see AI as the tipping point that will make cardiac CT more widely adopted as a first-line test for chest pain? #CaseTuesday #CardiacCT #AIinHealthcare #Radiology #HeartHealth #GEHealthcare

This paper evaluates an AI model for direct-to-physician reporting of ambulatory ECGs, comparing its performance to human technicians in detecting critical arrhythmias. 1️⃣ AI analyzed 14,606 ambulatory ECG recordings, with annotations from 167 certified technicians and 17 cardiologist panels. 2️⃣ AI demonstrated a 98.6% sensitivity for detecting critical arrhythmias, compared to 80.3% for human technicians. 3️⃣ False-negative findings were significantly lower for AI (3.2 per 1,000 patients) versus technicians (44.3 per 1,000 patients), making missed diagnoses 14.1 times more likely for technicians. 4️⃣ AI had a higher false-positive rate (12 per 1,000 patient days) than technicians (5 per 1,000 patient days). 5️⃣ AI showed a 99.9% negative predictive value, meaning it reliably ruled out critical arrhythmias with minimal risk of missing true cases. 6️⃣ AI performed consistently across gender groups and showed increasing accuracy with longer ECG monitoring durations. 7️⃣ Direct-to-physician AI reporting could reduce costs, improve access to ECG monitoring, and streamline healthcare workflow. ✍🏻 L.S. Johnson, Piotr Zadrożniak, G. Jasina, Agnieszka Grotek-Cuprjak, Jason Andrade, Emma Svennberg, Søren Zöga Diederichsen, W.F. McIntyre, Stavros (HSC) Stavrakis, J. Benezet-Mazuecos, Philipp Krisai, Zaza Iakobishvili, AVISHAG LAISH-FARKASH, S. Bhavnani, E. Ljungström, J. Bacevicius, N.L. van Vreeswijk, Michiel Rienstra, Raphael Spittler, J.A. Marx, Alireza Oraii, A. Miracle Blanco, A. Lozano, I. Mustafina, J. Healey, et al. Artificial intelligence for direct-to-physician reporting of ambulatory electrocardiography. Nature Medicine. 2025. DOI: 10.1038/s41591-025-03516-x

From Acquisition to Analysis: How AI is Revolutionizing Cardiac MRI by Solenn Toupin, Ph.D. and Théo Pezel, M.D., Ph.D. (Lariboisière Hospital, MIRACL.ai, Multimodality Imaging for Research and Analysis Core Laboratory: Artificial Intelligence, AP-HP, Paris, France). Artificial intelligence (#AI) is emerging as a powerful ally in cardiac #MRI, addressing many of the challenges that previously limited its efficiency and accessibility. By automating and optimizing steps from protocol planning and image acquisition to reconstruction, analysis, and integration with clinical data, AI can make cardiac MRI faster, more consistent, and more widely available. Far from replacing clinicians, AI supports them by reducing repetitive tasks, improving reproducibility, and enabling the extraction of advanced diagnostic and prognostic information. An important aspect of this evolution is the integration of cardiac MRI into a multimodality framework where it is combined with other imaging techniques such as echocardiography or CT, and with clinical, biological, and electrophysiological data. This approach paves the way for advanced concepts like the digital twin – a virtual model of the patient’s heart that can guide diagnosis and therapy planning, further enhancing precision and personalization in cardiovascular care. The authors explore how AI is transforming their cardiac MRI practice in four main domains: 1. Planning and acquisition: including automated plane prescription and parameter optimization 2. Image reconstruction: accelerating acquisitions and improving image quality 3. Image analysis and post-processing: enabling rapid and consistent quantification 4. Development of diagnostic and prognostic tools: integrating imaging with multisource and multimodal patient data Continue reading: https://lnkd.in/di6k3PED #MagnetomWorld #WhyCMR #CardiacMRI Gaia Banks Siemens Healthineers 

The Future of Cardiology: How AI is Transforming ECG Analysis with LLMs Imagine if AI could act like a virtual cardiologist, interpreting complex ECG data with precision, speed, and reliability. Thanks to recent breakthroughs, this future is closer than we think, as AI evolves to bridge the gap between raw medical data and actionable insights. 🔹 Research Focus A groundbreaking study by the authors explores integrating Large Language Models (LLMs) with electrocardiogram (ECG) data to enhance diagnostic accuracy. Unlike traditional methods, which convert ECG signals into simplified text, this approach leverages detailed ECG embeddings to maximize data fidelity and reasoning capabilities. 🔹 Challenges in Previous Methods Traditional techniques compressed ECG information into basic summaries using external classifiers. While convenient, this reduced the potential for nuanced AI-driven analysis, as much of the valuable signal detail was lost. Furthermore, the field lacked mature pre-trained ECG encoders, a gap addressed by the proposed model. 🔹 Innovative Solution The authors developed a model that processes ECG embeddings directly via a novel projection layer, trained on 800,000 ECG-report pairs. By feeding these embeddings into an LLM, the system retains comprehensive ECG data, allowing for richer and more accurate interpretations. It also supports time-based comparisons, a key requirement in clinical settings. 🔹 Addressing Hidden Biases The study identified a confounder, “severity of illness”, which created spurious correlations between question phrasing and answers. For instance, specific questions might prompt answers unrelated to ECG data. Using a de-biasing pre-training method based on causal theory, the authors eliminated these biases, ensuring the model relies solely on ECG input for its reasoning. 🔹 Validation and Real-World Applications Testing on the ECG-QA dataset revealed exceptional robustness, outperforming traditional and previous LLM-based approaches, even in adversarial scenarios. The model’s zero-shot capabilities demonstrated its versatility across unseen datasets. Additionally, its ability to compare sequential ECGs provides critical insights for patient care, aiding in early detection of subtle changes. 📌 Key Takeaways This study represents a major step forward in AI-driven cardiology. By preserving ECG data integrity and addressing biases, the model delivers reliable, real-time diagnostics. Its implications extend beyond clinical utility, offering strategic opportunities for businesses investing in healthcare innovation and technology. 👉 How do you see AI shaping the future of cardiology? What barriers might exist to implementing AI-driven ECG analysis in clinical practice? 👈 #ArtificialIntelligence #GenerativeAI #FutureOfWork #HealthTech #HealthcareInnovation

The headline that caught my eye this week was "AI Trial to Spot Heart Condition Before Symptoms." Here's my take: Artificial intelligence holds substantial promise to improve quality and reduce costs in healthcare. One example from Leeds involves an algorithm that scours medical records for early warning signs of atrial fibrillation (AF) before symptoms appear — potentially preventing thousands of strokes. The results suggest that by analyzing existing medical records for patterns that human physicians might miss, AI can flag high-risk patients for early intervention. The trial has already identified cases like a 74-year-old former Army captain who had no symptoms but can now manage his condition effectively. This is particularly significant given that AF contributes to around 20,000 strokes annually in the UK alone. As Professor Chris Gale notes, too often the first sign of undiagnosed AF is a stroke — an outcome this technology could help prevent. The broader implication here is about AI's role in healthcare: not replacing physicians but augmenting their ability to identify risks earlier and intervene before conditions become critical.  

Heterogeneous datasets are pervasive today, existing in various domains. Objects within these complex datasets are often represented from different perspectives, at different scales, or through multiple modalities, such as images, sensor readings, language sequences, and compact mathematical statements. Such datasets have been analyzed in the past using Multi-View Learning (MVL), Multi-Task Learning (MTL), and Tensor Learning (TL). In recent years, Multi-Modal Learning (MML) has also been employed. MML is a Machine Learning (ML) approach that integrates and processes information from multiple types of data, with different "perspectives" or "modalities" such as text, images, audio, video, or sensor data. The goal of MML is to leverage the complementary strengths of these modalities to improve model performance and enable richer understanding and predictions. Precision medicine and personalized clinical decision support systems (CDSS) tools have long aimed to leverage multimodal patient data to better capture complex, high-dimensional patient states and provider responses. This data ranges from free-form text notes and semi-structured electronic health records (EHR) to high-frequency physiological signals. While the advent of transformer architectures has enabled deeper insights from merging modalities, it has also required meticulous feature engineering and alignment. In patient monitoring, effectively analyzing diverse physiological signals within CDSS is highly challenging. #MedicalInformatics To address the challenges of analyzing multimodal patient data, the authors of [1] introduce MedTsLLM, a general multimodal large language model (LLM) framework that effectively integrates time series data and rich contextual information in the form of text. This framework performs three clinically relevant tasks (in time-series) which enable deeper analysis of physiological signals and can provide actionable insights for clinicians: • semantic segmentation • boundary detection • anomaly detection At a high level, boundary detection splits signals into periods like breaths or beats. Semantic segmentation further splits time series into distinct, meaningful segments. Anomaly detection identifies periods within the signals that deviate from normal. MedTsLLM utilizes a reprogramming layer to align embeddings of time series patches with a pretrained LLM's embedding space, making effective use of raw time series in conjunction with textual context. They additionally tailored the text prompt to include patient-specific information. Their experiments showed that MedTsLLM outperforms state-of-the-art baselines, including deep learning models, other LLMs, and clinical methods, across multiple medical domains, specifically electrocardiograms (ECG) and respiratory waveforms. Links to their preprint [1] and #Python GitHub repository [2] are shared in the comments.

AI Spots a Rare Heart Killer from a Single Echo—Now with FDA Breakthrough Status Cardiac amyloidosis is often missed until it’s too late. A new AI-enhanced echocardiography tool can flag it from one standard echo video clip—and it just received FDA Breakthrough Device designation. Why this matters: First commercially available AI echo tool to screen for amyloid cardiomyopathy Performance: AUROC 0.93, 85% sensitivity, 93% specificity, 96% NPV Impact: Earlier detection → earlier treatment → better outcomes for a notoriously underdiagnosed cause of heart failure Clinician takeaways: Consider AI screening in patients with unexplained LV wall thickening, HFpEF features, or overlapping phenotypes (e.g., HTN, HCM, AS) One clip, minimal workflow friction—ideal for broad screening in echo labs I’ve long advocated for the shift from reactive to proactive medicine. This is that future—arriving in the echo suite. Would you pilot this in your lab? What safeguards or workflows would you want in place first? (Link in comments.) #AIinHealthcare #Cardiology #Echocardiography #DigitalHealth #MedicalAI #Amyloidosis #HeartFailure #EarlyDiagnosis #HealthTech #FDA #MayoClinic

European Commission: #AI-driven Innovation in #Medical #Imaging Artificial Intelligence (AI) is nowadays increasingly used in the healthcare sector to support the prevention, early diagnosis, monitoring, and outcome prediction of non-communicable diseases (NCDs), which account for most deaths and disease burden across Europe. By automating tasks such as organ segmentation, lesion detection, disease classification, and uncertainty quantification, AI technologies offer the potential to reduce clinician workload, improve diagnostic accuracy, and enable earlier interventions. This report reviews the most widely adopted AI methods in medical imaging, focusing on segmentation, detection, and classification tasks, and assesses their technological maturity and readiness for clinical uptake. It illustrates key technical and systemic requirements through two practical use cases. The first use case is an end-to-end AI pipeline for lung cancer imaging, which includes full chest multi-organ segmentation from CT scans, pulmonary nodule detection and segmentation, longitudinal tracking of nodules, and uncertainty quantification. To support large-scale, privacy-preserving research and clinical deployment, this use case also develops a complete workflow for data anonymisation, secure upload, and web-based visualisation and validation by clinicians. The second use case focuses on cardiovascular disease classification, leveraging a biomechanics-informed model that extracts physiologically meaningful features from cine cardiac MRI sequences to enhance explainability and trustworthiness. The findings demonstrate that while AI holds strong promise to transform healthcare delivery for NCDs, achieving clinical deployment requires careful attention to data access, model transparency, and robust validation. Based on technical outcomes and operational experience, the study offers concrete recommendations to support EU initiatives aiming at the safe, effective, and trustworthy development of AI in healthcare.