Weakly Semi-supervised phenotyping using Electronic Health records.

Area of Science:

• Biomedical informatics
• Machine learning in healthcare
• Electronic Health Record (EHR) data analysis

Background:

• Electronic Health Record (EHR) phenotyping is critical but challenging due to data volume and heterogeneity.
• Manual chart review is time-consuming and expensive, leading to a scarcity of clinical annotations.
• Existing weakly supervised methods struggle with optimal classifier cutoff selection and may miss episodic phenotypes.

Purpose of the Study:

• To propose a label-efficient, weakly semi-supervised deep learning algorithm (WSS-DL) for EHR phenotyping.
• To overcome limitations of existing weakly supervised methods in EHR phenotyping.
• To assess the generalizability and label efficiency of WSS-DL across diverse phenotypes and healthcare systems.

Main Methods:

• WSS-DL employs a multi-stage learning process: generating silver standard labels, deriving enhanced-silver-standard labels using a weakly supervised deep learning model, and final classification with a supervised model using minimal gold standard labels.
• The algorithm was applied to classify 17 diseases (acute and chronic) using EHR data from three healthcare systems.
• The study determined the minimum quantity of training labels required for WSS-DL to outperform existing methods.

Main Results:

• WSS-DL effectively leverages EHR features from unlabeled samples through deep learning's ability to handle high-dimensional data.
• The method generates strong phenotype status predictions from silver standard labels, which serve as effective features in the final classification stage.
• High phenotyping accuracy was achieved with notably small subsets of labeled data (e.g., 40 labeled samples).

Conclusions:

• The proposed WSS-DL method demonstrates high performance in EHR datasets with very limited labeled data.
• This approach shows significant potential for assisting clinicians in diagnosing rare diseases.
• The method is valuable for identifying conditions that are susceptible to misdiagnosis due to data limitations.