Prior Adaptive Semi-supervised Learning with Application to EHR Phenotyping.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Health Services Research

Background:

• Electronic Health Records (EHR) offer rich data for biomedical research but are underutilized due to imprecise phenotype information.
• Supervised learning methods for EHR phenotyping require large labeled datasets, which are often unavailable.
• Existing methods struggle with large feature sets and limited gold-standard labeled data.

Purpose of the Study:

• To develop a novel semi-supervised (SS) EHR phenotyping method to address the limitations of supervised approaches.
• To improve the accuracy and generalizability of phenotype prediction from EHR data.
• To leverage both small labeled and large weakly-labeled datasets for enhanced phenotyping.

Main Methods:

• Proposed a semi-supervised (SS) EHR phenotyping approach utilizing a small labeled dataset and a large weakly-labeled dataset.
• Introduced a prior adaptive semi-supervised (PASS) estimator that incorporates prior knowledge by shrinking towards a derived direction.
• Derived asymptotic theory to justify the estimator's efficiency and robustness, even with imperfect prior information.

Main Results:

• The proposed PASS estimator demonstrated superiority over existing methods in simulation studies.
• The method proved effective and robust across various scenarios, including those with poor quality prior information.
• Validation on three real-world EHR phenotyping studies at a major tertiary hospital confirmed its practical utility.

Conclusions:

• The developed semi-supervised (SS) EHR phenotyping method, particularly the PASS estimator, significantly enhances phenotype prediction accuracy.
• This approach effectively overcomes the limitations of small labeled datasets in EHR research.
• The findings suggest a promising direction for advancing discovery research using Electronic Health Record data.