Polar labeling: silver standard algorithm for training disease classifiers.

Area of Science:

• Computational biology
• Medical informatics
• Machine learning

Background:

• Expert-labeled data are crucial for training phenotyping algorithms in cohort identification.
• Current expert labeling methods are time-consuming and expensive, limiting scalability.

Purpose of the Study:

• To introduce Polar Labeling (PL) as an efficient method for creating silver standard datasets.
• To evaluate if machine learning models trained on PL-generated data perform comparably to those trained on expert-labeled gold standard data.

Main Methods:

• Developed and applied the Polar Labeling (PL) approach to unlabeled patient records.
• Trained machine learning models on both PL-generated silver standard and manually created gold standard datasets.
• Validated the approach using health records from 38,023 patients across six diseases.

Main Results:

• Machine learning models trained with Polar Labeling demonstrated performance comparable to models trained on gold standard data.
• The proposed PL approach showed superior performance in disease classification tasks.
• Experimental validation confirmed the effectiveness of PL across multiple diseases.

Conclusions:

• Polar Labeling provides a scalable and cost-effective alternative to expert labeling for training machine learning models in healthcare.
• This method facilitates the development of robust phenotyping algorithms for broader clinical applications.
• The study highlights the potential of PL to accelerate the use of AI in disease classification.