Constrained Tensor Factorization for Cancer Phenotyping and Mortality Prediction.

Area of Science:

• Computational biology
• Health informatics
• Machine learning in oncology

Background:

• Electronic health records (EHR) contain vast patient data.
• Machine learning (ML) methods can extract valuable information from EHR.
• Computational phenotypes derived from EHR can aid in clinical prediction.

Purpose of the Study:

• To apply tensor factorization to EHR data for predicting five-year cancer mortality.
• To evaluate the impact of supervised terms, indication filtering, and social determinants of health (SDOH) on predictive model performance.
• To assess the effectiveness of constrained tensor factorization for phenotype extraction from sparse EHR data.

Main Methods:

• Utilized Northwestern Medicine EHR data from 2000-2015.
• Analyzed cohorts for breast, prostate, colorectal, and lung cancer.
• Employed constrained tensor factorization with added supervised terms, indication filtering, and SDOH covariates.

Main Results:

• Model performance varied across cancer types, with AUCs ranging from 0.517 to 0.750.
• Improvements in interpretability and performance were observed with the addition of supervised terms, indication filtering, and SDOH covariates.
• Constrained tensor factorization demonstrated effectiveness in deriving mortality-predictive phenotypes.

Conclusions:

• Constrained tensor factorization is a viable method for extracting mortality-predictive phenotypes from sparse EHR data.
• Incorporating SDOH and supervised learning techniques enhances the performance and interpretability of EHR-based predictive models.
• This approach offers a pathway for improved cancer patient outcome prediction using routinely collected health data.