Cancer patients undergoing activity treatment often face adverse clinical events, including unexpected re-hospitalizations and mortality. Currently, doctors must manually interpret vast amounts of clinical and genomic data in order to assess a proper treatment plan for a patient. In order to leverage the capabilities of this data, we aim to use machine learning models to predict patients that have high readmission and mortality risks and specifically what features or variables are important for building such a model. That way, clinicians can be informed of when to make the necessary population health interventions. We are using real-world EHR and genomic data from the UC San Diego Moores Cancer Center.

We initially performed feature processing and literature review to extract features from the clinicogenomic data that are relevant to hospitalization events. To conduct this feature selection, we used multiple models for the two types of data we extracted from our feature selection: categorial and continuous data. We used logistic and linear regression for our continuous data and CART decision trees and random forests for our categorical data. We were successfully able to extract features from various variable subsets that included Disease states, Medications, Encounter Types,  Social History Types, and Mutations that have correlations with predicting rehospitalization within 30 days for various cancer types. Our future directions include creating a comprehensive model with the features collected to verify the validity of the extracted features.