Learning endometriosis phenotypes from patient-generated data.

Area of Science:

• Computational biology
• Digital health
• Women's health

Background:

• Endometriosis is a chronic, enigmatic condition affecting women of childbearing age.
• It lacks known biomarkers and established clinical staging, hindering effective management.
• Patient-generated health data offers a novel avenue for understanding disease heterogeneity.

Purpose of the Study:

• To characterize endometriosis patient subtypes using unsupervised learning.
• To leverage self-tracked health data from smartphones for data-driven phenotyping.
• To identify clinically relevant endometriosis phenotypes from multimodal patient data.

Main Methods:

• Utilized self-tracking data from over 4000 women with endometriosis over 2 years.
• Extended a mixed-membership model to handle multimodal and uncertain self-tracked variables.
• Employed unsupervised learning to identify patient subtypes based on symptoms, quality of life, and treatments.

Main Results:

• Identified clinically relevant endometriosis subtypes through data-driven phenotyping.
• The proposed method demonstrated robustness to data biases and hyperparameter variations.
• Learned phenotypes align with existing endometriosis knowledge and suggest new actionable findings.

Conclusions:

• Unsupervised learning of endometriosis subtypes from self-tracked data shows significant promise.
• Patient-generated health data and digital technologies can advance the study of enigmatic diseases.
• This approach offers a path toward improved understanding and management of endometriosis.