🤖 machine learning

Dynamical Predictive Modelling of Cardiovascular Disease Progression Post-Myocardial Infarction via ECG-Trained Artificial Intelligence Model

This paper proposes a foundation model that leverages self-supervised contrastive learning on unlabelled ECGs combined with supervised multitask fine-tuning to significantly improve the prediction of post-myocardial infarction outcomes in data-scarce clinical settings.

Original authors: Riccardo Cavarra, Lupo Lovatelli, Shaheim Ogbomo-Harmitt, Shahid Aziz, Adelaide De Vecchi, Andrew King, Oleg Aslanidi

Published 2026-05-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Riccardo Cavarra, Lupo Lovatelli, Shaheim Ogbomo-Harmitt, Shahid Aziz, Adelaide De Vecchi, Andrew King, Oleg Aslanidi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your heart as a complex orchestra. Every time it beats, it plays a unique song captured by a machine called an ECG (electrocardiogram). For decades, doctors have used these recordings to spot immediate problems, like a heart attack happening right now. But once the heart attack is over, doctors often stop listening to the "music" of the ECG, even though the song might hold clues about whether the patient will get sick again later.

The problem is that teaching computers to understand these songs is hard. Usually, you need a massive library of songs where every single one is labeled with a specific diagnosis (like "this song means heart failure") to teach the computer. In medicine, these labeled libraries are rare and small. When you try to teach a computer with too few examples, it gets confused and starts guessing based on silly shortcuts rather than learning the real rules of the music.

The Solution: A "Musical Ear" That Learns on Its Own

The researchers in this paper built a new kind of Artificial Intelligence (AI) that learns differently. Instead of waiting for a teacher to label every song, they let the AI listen to 800,000 unlabelled ECG recordings from a massive database.

Think of this AI as a student with a superpower: Pattern Recognition.

The "Same Person" Game (Contrastive Learning):
The AI was taught a simple game: "If you hear two songs recorded from the same person within a 60-day window, they are related, even if they sound slightly different due to background noise or how the machine was held."
- The Analogy: Imagine you are trying to recognize a friend's voice. You might hear them whispering, shouting, or singing off-key. Even though the sound changes, you know it's still them. The AI learned to ignore the "noise" (like a cough or a shaky hand) and focus on the unique "voice" of the patient's heart.
- The Twist: The AI was also told that songs from different people, or the same person recorded years apart, are completely unrelated. This helped the AI understand how a heart's "song" changes over time as a disease progresses.
The "Music Theory" Class (Supervised Tasks):
While the AI was playing the "Same Person" game, the researchers also gave it a few specific homework assignments to keep it on track. They asked it to identify specific musical notes (like irregular heart rhythms) and measure the length of specific beats (like the time between heartbeats).
- The Analogy: This is like a music student who practices improvisation (the unlabelled game) but also takes a few theory classes to ensure they understand the basics of rhythm and pitch. This prevents the AI from getting lost in the noise and ensures it learns medically useful information.

The Big Test: Predicting the Future

Once the AI had spent months "listening" to 800,000 heartbeats and learning the structure of heart music, the researchers tested it on a new, smaller group of patients who had recently had a heart attack. They asked the AI two questions:

Will this patient die?
Will this patient develop heart failure?

They compared their "smart" AI (which had listened to the massive library first) against a "dumb" AI (a standard computer model that started with zero knowledge and only looked at the small group of patients).

The Results

The difference was like comparing a seasoned music critic to someone who has never heard a symphony before.

The "Dumb" AI: Scored poorly (around 0.60 out of 1.0). It struggled to find the signal in the noise because it didn't have enough examples to learn from.
The "Smart" AI: Scored much higher (around 0.79 out of 1.0). Because it had already learned the "grammar" of heart signals from the massive library, it could spot the subtle signs of future trouble in the small group of patients.

The Bottom Line

The paper concludes that by teaching an AI to understand the temporal story of a patient's heart (how their specific heart rhythm evolves over time) using a huge amount of unlabelled data, we can build a much better tool for predicting who is at risk after a heart attack.

In short: Instead of trying to teach a computer with a tiny dictionary, the researchers gave it a massive library to read first. Now, when it looks at a new patient, it doesn't just see a squiggly line; it hears the story of a heart that is struggling, allowing it to predict future risks much more accurately than before.

Technical Summary: Dynamical Predictive Modelling of Cardiovascular Disease Progression Post-Myocardial Infarction via ECG Pre-Trained AI Model

Problem Statement
Myocardial infarction (MI) remains a leading cause of mortality and long-term morbidity. While the electrocardiogram (ECG) is a standard, non-invasive tool for initial MI stratification, it currently plays no role in post-MI patient management. The authors identify a critical gap: deep learning (DL) models for ECG-based prognosis underperform in clinical settings because they rely on large, labelled datasets which are scarce in medicine. In limited data regimes, standard DL strategies often fail to extract clinically relevant information, instead relying on "shortcuts" that lead to poor generalization. Furthermore, existing foundation models utilizing self-supervised learning often fail to incorporate patient-level temporal context, which is essential for capturing disease progression.

Methodology
The proposed solution is a pre-trained artificial intelligence model that learns patient-specific, dynamics-aware representations by combining contrastive learning with supervised multitask objectives.

Data Sources: The model was pre-trained on 800,035 12-lead ECG recordings from 161,352 unique patients sourced from the MIMIC-IV database. For downstream evaluation, two specific post-MI cohorts were utilized: a mortality cohort (8,499 patients, 3,097 deaths) and a heart failure cohort (852 patients, 552 events).
Pre-training Strategy:
- Contrastive Learning: The core of the pre-training is a self-supervised approach where pairs of ECG signals from the same patient, recorded within a 60-day temporal window ( $T_w$ ), are treated as positive pairs. Signals from different patients or outside this window are treated as negative pairs. This forces the model to learn representations sensitive to disease-relevant changes over time while remaining invariant to acquisition noise.
- Augmentation: Stochastic data augmentations (Gaussian noise, amplitude scaling, time shifts, lead dropout, etc.) are applied to create distorted views of the same signal, preventing shortcut learning.
- Architecture: A 1D convolutional encoder (ResNet1D) processes the augmented signals, followed by a projection head. The model optimizes a Normalized Temperature-scaled Cross Entropy (NT-Xent) loss to maximize similarity between positive pairs.
- Multitask Supervision: To ensure clinical relevance, the encoder is simultaneously trained with linear task heads for rhythm classification (atrial fibrillation, ST elevation) and waveform regression (P-wave, QRS, T-wave, RR intervals). These supervised tasks act as regularizers, weighted by learned homoscedastic uncertainty to handle different task scales.
Fine-tuning: The pre-trained ResNet1D encoder is transferred to downstream tasks. The encoder weights are frozen or fine-tuned, and a new dense layer network is trained to classify mortality and heart failure using single post-MI ECG instances.

Key Contributions

Dynamical Contrastive Learning: The paper reframes contrastive learning as a dynamical modelling problem by defining positive pairs based on patient identity and a clinically meaningful time window (60 days), rather than just augmented views of a single instance.
Hybrid Training Objective: The integration of self-supervised contrastive learning with supervised multitask heads (rhythm and morphology) ensures the learned embeddings capture both patient-specific temporal dynamics and clinically structured features.
Performance in Low-Data Regimes: The study demonstrates that this pre-training strategy significantly improves prognostic performance on small, labelled post-MI datasets compared to training from scratch.

Results
The study evaluated the pre-trained model against a baseline ResNet1D trained from scratch on the same downstream tasks. Performance was measured using the Area Under the Receiver Operating Characteristic Curve (AUC):

Mortality Prediction: The pre-trained model achieved an AUC of 0.794, compared to 0.608 for the model trained from scratch (a 19% improvement).
Heart Failure Prediction: The pre-trained model achieved an AUC of 0.781, compared to 0.602 for the baseline (an 18% improvement).

Significance and Claims
The authors claim that this work serves as a proof-of-concept for using "clinically smart" pre-trained AI models in low-data regimes typical of cardiovascular outcome prediction. They assert that by structuring ECG modelling to account for patient-level temporal context and disease progression, the model learns representations that are more robust and clinically meaningful than those derived from scratch. The results suggest that such representation learning strategies can effectively improve long-term outcome prediction for post-MI patients, potentially enabling the use of ECGs for ongoing patient management where they are currently limited.

The Solution: A "Musical Ear" That Learns on Its Own

The Big Test: Predicting the Future

The Results

The Bottom Line

Technical Summary: Dynamical Predictive Modelling of Cardiovascular Disease Progression Post-Myocardial Infarction via ECG Pre-Trained AI Model

More like this