Cardiac Foundation Model Using ECG-PCG Signals With Less Labeled Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing model training protocols for cardiovascular conditions require large amounts of labeled data, which is time-consuming and often biased due to the difficulty in obtaining ground truth conditions, especially for rare conditions, leading to challenges in training effective models.
Innovation Solution
A method involving a foundation model trained with synchronized electrocardiogram (ECG) and phonocardiogram (PCG) data, utilizing unlabeled data to learn associations between features, and fine-tuning with a smaller set of labeled data to generate specialized models for conditions like atrial fibrillation, murmur, and pulmonary hypertension.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning models are trained with large amounts of labeled training data, then model accuracy may improve, but the time consumption and complexity of data labeling increase significantly
Solution Approach 1:
The patent applies preliminary action by pre-training a foundation model on large volumes of unlabeled ECG and PCG data using self-supervised learning before fine-tuning on smaller labeled datasets. This preliminary training phase enables the model to learn fundamental patterns and associations from abundant unlabeled data, reducing the subsequent need for extensive manual labeling while maintaining high accuracy.
Solution Approach 2:
The patent introduces a foundation model as an intermediary between unlabeled data and the final specialized model. This foundation model serves as a mediator that first processes unlabeled data to learn general cardiovascular patterns, then guides the fine-tuning process on labeled data, thereby reducing the direct burden of labeling while achieving accurate specialized models.
2Measurement precision
If traditional machine learning models are trained with labeled data confirmed by invasive procedures, then ground truth accuracy improves, but the difficulty and risk of data collection increase
Solution Approach 1:
The patent replaces expensive and invasive ground truth verification methods with cheaper, non-invasive alternatives. By using the foundation model's predictions and less invasive clinical assessments as surrogate labels, the system avoids the need for complex invasive procedures while still achieving sufficient training data quality for effective model fine-tuning.
Solution Approach 2:
The patent enables the training system to partially self-service by using the foundation model's own predictions and unsupervised learning capabilities to generate training signals. The model learns to identify patterns and associations without requiring external invasive verification for every data point, reducing the overall complexity of data collection while maintaining training effectiveness.
3Measurement precision
If labeled training data is assembled from rare conditions in the overall patient population, then model specificity for rare conditions improves, but the quantity of available training data decreases
Solution Approach 1:
The patent applies preliminary action by first training the foundation model on large volumes of common condition data to learn general cardiovascular patterns and associations. This preliminary training on abundant data establishes a strong base that can then be efficiently fine-tuned on smaller datasets of rare conditions, improving the model's ability to detect rare conditions even with limited specific training examples.
Solution Approach 2:
The patent creates a universal foundation model that learns general cardiovascular patterns from diverse conditions, making it applicable to multiple specific conditions including rare ones. This multi-functional foundation model can be fine-tuned for different conditions (atrial fibrillation, murmur, pulmonary hypertension, etc.) using smaller condition-specific datasets, thereby improving rarity detection without requiring large condition-specific datasets.
Data Source
Figure 1A
Figure 1B
Figure 2
AI summary
The present description relates generally to methods and systems for detecting cardiovascular conditions using a foundation model. In one example, a method includes obtaining a synchronized ECG signal and a PCG signal from a patient, converting the PCG signal to a PCG mel-spectrogram, entering the ECG signal and the PCG mel-spectrogram as input to a trained specialized model configured to output a classification output based on the ECG signal and the PCG mel-spectrogram, the trained specialized model trained with labeled ECG and PCG signal pairs using a foundation model trained with unlabeled ECG and PCG signal pairs, and storing the classification output in memory and/or displaying the classification output on a display device.