Scalable Data-driven Phenotypes via Unsupervised Feature Learning.

Area of Science:

• Computational biology
• Bioinformatics
• Medical informatics

Background:

• Personalized medicine relies on identifying patient phenotypes from clinical data.
• Current methods use supervised learning, which requires expert input and can miss novel patterns.

Purpose of the Study:

• To introduce unsupervised feature learning for phenotype recognition.
• To develop a method for inferring data-driven micro-phenotypes from electronic medical record (EMR) data.

Main Methods:

• Applied unsupervised feature learning to longitudinal laboratory values from EMR data.
• Evaluated the learned features in a classification task without prior knowledge of labels.

Main Results:

• Unsupervised features were learned from noisy, sparse, and irregular EMR data.
• The learned features achieved an area under the curve (AUC) of 0.96 in a classification task.
• Performance was comparable to expert-engineered features.

Conclusions:

• Unsupervised feature learning is a viable and effective approach for phenotype recognition.
• This method can uncover broadly useful, data-driven micro-phenotypes.
• It offers a scalable alternative to supervised methods for analyzing complex clinical data.