A deep foundation model for electrocardiogram interpretation: enabling rare disease detection through transfer...

Area of Science:

• Artificial Intelligence in Medicine
• Cardiology
• Machine Learning for Healthcare

Background:

• Deep neural networks (DNNs) in healthcare are limited by scarce high-quality labeled data.
• Foundation models offer an efficient deep learning approach, enabling effective DNN training with reduced labeled examples.
• Electrocardiograms (ECGs) are crucial for diagnosing cardiac conditions, but comprehensive AI models are needed.

Purpose of the Study:

• To pre-train a comprehensive deep neural network (DNN) foundation model using a large dataset of electrocardiograms (ECGs).
• To evaluate the utility of this ECG foundation model for detecting novel diagnoses with limited data through fine-tuning.
• To compare the performance of fine-tuned models against models trained from scratch on small datasets.

Main Methods:

• Leveraged cardiologist-confirmed labels from 1.6 million ECGs (1986-2019) to pre-train a convolutional DNN.
• Assessed the pre-trained model's performance on 68 common ECG diagnoses (median AUC 0.978).
• Fine-tuned the foundation model on small datasets to detect carcinoid syndrome, pericardial constriction, and rheumatic mitral valve doming.

Main Results:

• The comprehensive ECG DNN foundation model achieved high performance (median AUC 0.978, sensitivity 0.937, specificity 0.923).
• Fine-tuning the foundation model significantly improved detection of novel ECG diagnoses compared to training from scratch.
• Achieved AUCs of 0.772 (carcinoid syndrome), 0.883 (pericardial constriction), and 0.826 (rheumatic doming) via fine-tuning.

Conclusions:

• The developed ECG foundation model learns flexible ECG waveform representations.
• Foundation models enhance downstream model performance, particularly in data-limited scenarios for rare conditions.
• This approach facilitates the development of robust AI tools for complex cardiac diagnostics.