Semi-supervised Learning for Phenotyping Tasks.

Area of Science:

• Computational biology
• Medical informatics
• Machine learning

Background:

• Supervised learning is standard for electronic health records (EHR)-based phenotyping but requires expensive manual chart review.
• Semi-supervised learning leverages both limited labeled and abundant unlabeled data to improve efficiency.

Purpose of the Study:

• To investigate a family of semi-supervised learning algorithms based on Expectation Maximization (EM) for EHR phenotyping.
• To evaluate the effectiveness of basic and Augmented EM algorithms in reducing the need for manual data labeling.

Main Methods:

• Exploration of Expectation Maximization (EM) and Augmented EM algorithms for phenotyping tasks.
• Assessment of weighting factors in Augmented EM, considering cross-validation and heuristic methods.
• Training phenotyping models with limited labeled and extensive unlabeled EHR data.

Main Results:

• Basic EM can yield inaccurate parameter estimates when modeling assumptions are unmet.
• Augmented EM, with a weighting factor, mitigates inaccuracies by downweighting unlabeled data.
• Accurate phenotyping models were achieved using only a few hundred labeled examples alongside numerous unlabeled ones.

Conclusions:

• Semi-supervised learning, particularly Augmented EM, offers a viable and cost-effective solution for EHR phenotyping.
• This approach can substantially decrease the manual chart review burden, enabling significant cost savings.
• Careful selection of weighting factors, potentially using heuristic methods over cross-validation, is crucial for optimal performance.