TASTE: Temporal and Static Tensor Factorization for Phenotyping Electronic Health Records.

Area of Science:

• Computational Medicine
• Biomedical Informatics
• Data Science

Background:

• Electronic Health Records (EHR) phenotyping aims to identify patient cohorts and their disease progression.
• Existing tensor factorization methods often model static or temporal EHR data separately, failing to capture the complexity of combined data types.
• Simultaneously modeling both static (demographics) and temporal (visits) EHR data presents a significant challenge.

Purpose of the Study:

• To propose Temporal And Static TEnsor factorization (TASTE), a novel method for joint modeling of static and temporal EHR data for improved phenotyping.
• To develop an efficient computational framework for fitting the proposed TASTE model.
• To evaluate the clinical meaningfulness and predictive performance of phenotypes extracted by TASTE.

Main Methods:

• Developed TASTE, combining PARAFAC2 and non-negative matrix factorization to model static and temporal tensors from EHR data.
• Transformed the model fitting into optimally solved sub-problems using alternating optimization.
• Implemented efficient sub-problem solvers through novel mathematical re-formulations.

Main Results:

• TASTE demonstrated up to 14x speed improvement over baseline methods on large-scale EHR data.
• Phenotypes extracted by TASTE were validated as clinically meaningful by a cardiologist.
• A logistic regression model using 60 TASTE-derived phenotypes achieved comparable Area Under the Curve (AUC) to a deep learning model with 345 features for heart failure prediction.

Conclusions:

• TASTE effectively integrates static and temporal EHR data for robust patient phenotyping.
• The proposed method offers significant computational efficiency and clinical relevance.
• TASTE-derived phenotypes enhance predictive modeling accuracy for diseases like heart failure.