ECG Foundation Models and Medical LLMs for Agentic Cardiovascular Intelligence at the Edge: A Review and Outlook

This survey reviews the synergistic integration of ECG foundation models and medical large language models to develop agentic, edge-deployable cardiovascular intelligence systems that enable real-time, privacy-preserving diagnosis, monitoring, and clinical decision support on consumer-grade wearable devices.

The Gist

Imagine your heart is like a busy, 24/7 orchestra playing a complex symphony. For decades, doctors have been the conductors, listening to this music (your ECG) to spot a missed beat or a wrong note. But there are too many concerts happening every second across the world, and not enough conductors to listen to them all.

This paper is a roadmap for building a super-smart, tireless digital assistant that can listen to your heart's music, understand the lyrics of your medical history, and give you advice—all while sitting right on your wristwatch, without needing to call a giant server in the cloud.

Here is the breakdown of how this "Agentic Cardiovascular Intelligence" works, using simple analogies:

1. The Two Brains: The Ear and The Librarian

The paper argues that to build the perfect heart doctor, we need to combine two different types of AI "brains" that work together:

• The "Super Ear" (ECG Foundation Models):
  Imagine a musician who has listened to millions of heartbeats. They haven't been taught specific rules; they just know what a healthy rhythm sounds like versus a chaotic one. This is the ECG Foundation Model. It looks at the raw squiggly lines of your heart signal and instantly spots patterns (like "Oh, that looks like an irregular heartbeat"). It's the expert signal interpreter.
• The "Super Librarian" (Medical LLMs):
  Now, imagine a doctor who has read every medical textbook, research paper, and patient guideline ever written. This is the Medical Large Language Model (LLM). It doesn't listen to the heart directly; instead, it knows the rules. It knows that if a patient has high blood pressure and is taking a specific drug, a certain heartbeat pattern might be dangerous. It provides the context and the reasoning.

The Magic: When you combine them, you get an Agent. The "Super Ear" hears the problem, and the "Super Librarian" immediately checks the rulebook, looks at your history, and says, "This isn't just a glitch; it's a warning sign for a specific condition. Here is what you should do."

2. The Problem: The "Heavy Suitcase"

Usually, these super-smart AIs are like giant, heavy suitcases. They are so big and powerful that they can only run in massive data centers (the Cloud).

• The Issue: If your smartwatch has to send your heart data to the cloud, wait for the answer, and send it back, it takes too long (latency). Plus, it drains your battery, and your private heart data is traveling over the internet, which feels risky.
• The Goal: We need to shrink these giant suitcases down to fit inside a tiny backpack (your watch) so the AI can work offline, instantly, and privately.

3. The Solution: "Model Optimization" (The Art of Packing)

The paper explains how scientists are shrinking these giant AI brains so they can fit on your watch without losing their smarts. They use three main tricks:

• Quantization (The "Low-Res" Trick): Imagine a 4K movie. It looks amazing but takes up a lot of space. If you convert it to a slightly lower quality (like 720p), it still looks great, but the file size shrinks massively. The AI does this with its math, using simpler numbers to save space.
• Pruning (The "Weeding" Trick): Imagine a garden with thousands of plants, but only a few are actually flowers. Pruning is cutting away the dead leaves and useless branches. The AI cuts out the parts of its brain that aren't needed for heart monitoring, making it lighter.
• Knowledge Distillation (The "Teacher-Student" Trick): Imagine a genius professor (the big Cloud AI) teaching a bright student (the small Watch AI). The professor doesn't just give the student the whole textbook; they teach them the key concepts and how to think. The student becomes small and fast but retains the most important wisdom.

4. The Future: The "Agentic" Watch

The paper envisions a future where your watch isn't just a passive recorder; it's an active agent.

• Right Now: Your watch says, "Your heart rate is high."
• The Future (Agentic AI): Your watch says, "I noticed your heart rhythm changed at 2:00 PM. You mentioned you were stressed earlier, and your history shows you have a family risk of arrhythmia. Based on the guidelines, I recommend sitting down and breathing for 5 minutes. If this continues for 10 more minutes, I will automatically call your doctor."

It doesn't just detect; it reasons, plans, and acts.

5. Why This Matters

• Speed: It happens instantly on your device. No waiting.
• Privacy: Your heart data never leaves your watch. It stays with you.
• Battery: Because the AI is so efficient, your watch won't die in an hour.
• Accessibility: This technology could bring expert-level heart monitoring to people in remote villages or those who can't afford a specialist, simply by using a cheap smartwatch.

In a Nutshell

This paper is a blueprint for building a heart-protecting superhero that lives on your wrist. It combines a musician's ear (to hear the heart) with a librarian's mind (to understand the rules) and shrinks them down using clever engineering tricks so they can run quietly and securely in your pocket, ready to save your life the moment something goes wrong.

Technical Summary

1. Problem Statement

Cardiovascular diseases (CVDs) are the leading cause of global mortality. While Electrocardiograms (ECGs) are the gold standard for diagnosis and monitoring, their clinical utility is constrained by:

• Reliance on Expert Interpretation: Manual analysis is labor-intensive, variable, and difficult to scale.
• Limitations of Current AI: Traditional deep learning models (CNNs, RNNs) are task-specific, lack contextual reasoning, and struggle with generalization across heterogeneous data.
• Deployment Barriers: Large-scale Foundation Models (FMs) and Large Language Models (LLMs) are computationally too heavy for real-time, privacy-preserving deployment on consumer-grade edge devices (e.g., smartwatches).
• Fragmented Research: Existing surveys treat ECG signal processing, medical LLMs, and edge AI as isolated domains, lacking a unified framework for integrating them into a cohesive clinical system.

The paper identifies the need for a unified, agentic paradigm that combines signal-level interpretation with knowledge-based reasoning, optimized for edge deployment.

2. Methodology

The authors conducted a structured literature review (2020–2026) across major databases (IEEE Xplore, PubMed, Scopus, Google Scholar) focusing on the intersection of:

• ECG Foundation Models: Self-supervised learning on large-scale unlabeled waveform data.
• Medical LLMs: Knowledge-based reasoning trained on biomedical text.
• Edge AI: Model compression, optimization, and hardware-software co-design.

The review synthesizes these domains to propose a synergistic architecture where ECG FMs act as "signal interpreters" and Medical LLMs act as "reasoning backbones," connected via multimodal alignment and optimized for edge constraints.

3. Key Contributions

The paper makes three primary contributions:

A. The "Agentic" Paradigm for Cardiovascular AI

The authors propose a unified system where two complementary model classes operate in tandem:

1. ECG Foundation Models (Signal Interpreters): Pre-trained on massive unlabeled ECG datasets using self-supervised learning (SSL) to extract rich electrophysiological representations. They handle raw waveform analysis, anomaly detection, and feature extraction.
2. Medical LLMs (Reasoning Backbones): Trained on biomedical text (EHRs, guidelines, literature) to provide contextual inference, guideline alignment, and clinical decision support. They do not process raw signals directly but interpret the outputs of ECG models within a clinical context.

• Synergy: This creates an "agentic" system capable of end-to-end intelligence: detecting a rhythm irregularity (FM) and then generating a differential diagnosis or triage recommendation based on patient history and guidelines (LLM).

B. Comprehensive Review of Architectures and Techniques

• ECG Foundation Models: Reviews models like ECGFounder, HuBERT-ECG, and ECG-FM utilizing masked signal modeling, contrastive learning (CLOCS, TA-PCLR), and Vision Transformers (HeartBEiT).
• Multimodal Integration: Examines hybrid architectures (e.g., ECG-LM, MERL, PULSE) that align ECG embeddings with language models for zero-shot classification, automated report generation, and multimodal reasoning.
• Edge Optimization: Details critical techniques for deploying these heavy models on wearables:
  • Quantization: (e.g., 2-bit/4-bit quantization, LoRA integration) to reduce memory.
  • Pruning & Distillation: Removing redundant parameters and transferring knowledge from large "teacher" models to small "student" models.
  • Small Language Models (SLMs): Utilizing efficient architectures like Phi-3-mini and TinyLlama for on-device reasoning.
  • TinyML: Deployment on microcontrollers (e.g., GAP9, ESP32) with energy consumption in the microjoule range.
  • Federated Learning: Enabling collaborative training without sharing raw patient data to preserve privacy.

C. Deployment-Aware Roadmap

The paper outlines a translational roadmap connecting algorithmic advances to practical implementation, addressing:

• Hardware-Software Co-design: Matching models to specific edge hardware (e.g., Apple Neural Engine, Qualcomm Snapdragon).
• Privacy & Security: Addressing vulnerabilities like model inversion, adversarial attacks, and data leakage in edge-cloud pipelines.
• Future Directions: Proposes Agentic RAG (Retrieval-Augmented Generation), State Space Models (SSMs/Mamba) for efficient long-sequence processing, and Physical AI for proactive intervention.

4. Key Results and Findings

• Performance of Foundation Models: Models like ECGFounder (trained on 10.77M ECGs) and HuBERT-ECG (9.1M ECGs) achieve cardiologist-level performance in zero-shot and few-shot scenarios, outperforming traditional supervised models in generalization.
• Multimodal Success: Hybrid models (e.g., MERL, PULSE) successfully bridge the gap between signals and text, enabling automated report generation and reducing hallucinations through clinical knowledge grounding.
• Edge Feasibility:
  • Compression: Hybrid compression (pruning + quantization + distillation) can achieve >100x compression ratios with minimal accuracy loss.
  • Efficiency: TinyML implementations (e.g., NanoHydra) achieve >94% accuracy on microcontrollers with energy consumption as low as 7.69 µJ per inference, enabling multi-year battery life on wearables.
  • SLMs: Compact models (e.g., Phi-3-mini) can perform competitive diagnostic reasoning on smartphone-class devices.
• Privacy: Federated learning combined with differential privacy and homomorphic encryption is shown to be a viable strategy for training models across distributed devices without compromising patient data sovereignty.

5. Significance

This survey is significant because it moves beyond isolated technical reviews to propose a holistic, agentic framework for the next generation of cardiovascular care.

• Clinical Impact: It envisions a shift from reactive, hospital-centric monitoring to continuous, proactive, and personalized care delivered directly to patients via wearables.
• Technological Bridge: It bridges the gap between high-capability AI (FMs/LLMs) and resource-constrained reality (Edge/IoMT), providing a clear path for deployment.
• Trust and Safety: By emphasizing privacy-preserving techniques (Federated Learning) and explainability (Agentic reasoning), it addresses critical barriers to clinical adoption.
• Future Outlook: The paper sets the stage for autonomous cardiovascular agents that can not only detect arrhythmias but also reason about patient context, retrieve guidelines, and suggest actionable interventions in real-time.

In summary, the paper argues that the future of cardiovascular AI lies not in larger models alone, but in the synergistic integration of signal interpreters and reasoning agents, optimized for the edge to deliver scalable, secure, and real-time health intelligence.