Detecting time-evolving phenotypic topics via tensor factorization on electronic health records: Cardiovascular...

Area of Science:

• Cardiovascular disease research
• Machine learning in healthcare
• Precision medicine

Background:

• Discovering disease subphenotypes aids in developing targeted diagnostics and treatments.
• Current methods often overlook temporal aspects, modeling diseases as static events.
• Understanding disease evolution is crucial for precise phenotyping and comprehending progression.

Purpose of the Study:

• To evaluate the benefits of an unsupervised, time-aware approach for dynamic phenotype discovery.
• To model diseases as evolving processes using longitudinal electronic health record (EHR) data.

Main Methods:

• Applied constrained non-negative tensor factorization to longitudinal EHR data of cardiovascular disease (CVD) patients.
• Identified phenotypic topics (subphenotypes) and their temporal patterns over 10 years pre-CVD diagnosis.
• Assessed subphenotype associations with CVD risk and compared subsequent myocardial infarction (MI) rates using survival analysis.

Main Results:

• Identified 14 distinct subphenotypes from a cohort of 12,380 CVD individuals.
• Found associations between subphenotypes (e.g., Vitamin D deficiency, depression, urinary infections) and CVD risk, unexplained by conventional factors.
• Demonstrated significantly different MI risks among the six most prevalent subphenotypes (p < 0.0001).

Conclusions:

• Tensor decomposition effectively models diseases as dynamic processes using longitudinal EHR data.
• This data-driven approach can help identify complex, chronic disease subphenotypes in precision medicine.
• The identified subphenotypes appear clinically meaningful and offer insights into disease progression.