Cross-Modal Data Programming Enables Rapid Medical Machine Learning.

Area of Science:

• Medical Machine Learning
• Data Programming
• Weak Supervision

Background:

• Developing clinically impactful machine learning models is hindered by a lack of labeled training data.
• Medical researchers often use weaker, noisier supervision sources, like text extractions for image classification.
• Combining diverse, potentially error-correlated information sources presents a key challenge in weak supervision.

Purpose of the Study:

• To propose and evaluate a novel technique for cross-modal weak supervision in medicine using data programming.
• To leverage data programming for combining information from auxiliary modalities (e.g., text) to train models on target modalities (e.g., images).
• To assess the performance improvement and time efficiency compared to traditional hand-labeling methods.

Main Methods:

• Applied data programming for cross-modal weak supervision in medical imaging tasks.
• Utilized unstructured text reports as a weaker source of supervision for image classification.
• Evaluated the approach on diverse clinical tasks against institution-scale, hand-labeled datasets.

Main Results:

• The proposed technique improved model performance by up to 6 points ROC-AUC over baseline methods.
• Models achieved performance within 1.75 points ROC-AUC of physician-years of hand labeling.
• Outperformed models supervised with physician-months of hand labeling by 10.25 points ROC-AUC, saving 96% of time.

Conclusions:

• Data programming offers a powerful method for cross-modal weak supervision in medicine.
• This approach significantly enhances the efficiency and effectiveness of developing clinical machine learning models.
• Weak supervision techniques like data programming can accelerate the development and deployment of clinically useful AI tools.