EHR-Based Risk Prediction for Kidney Cancer.

Abstract

Background and objectives: The U.S. Preventive Services Task Force (USPSTF) does not currently recommend routine screening for kidney cancer, even though approximately 14 390 people are expected to die from this disease in the United States in 2024. Individualized risk-based kidney cancer screening offers the potential to effectively detect cancer at an early stage and avoid unnecessarily screening the rest of the population who are at low risk. This study proposes electronic health records (EHR) risk evaluation for kidney cancer by examining a comprehensive set of medical history including diagnoses, comorbidities, viruses, and rare diseases.

Methods: The relevant medical history for predicting kidney cancer occurrence was identified from the analysis of All of Us data in three steps. First, a Systematized Nomenclature of Medicine (SNOMED) code binary indicator variable in EHR was set for the presence of kidney cancer. Second, the relationship between this binary indicator of cancer and all prior health conditions was examined using the Strong Rule for Feature Elimination and Least Absolute Shrinkage and Selection Operator logistic regression methods of variable selection. Third, the accuracy of the model was reported using cross-validated McFadden's R2 and Area under the Receiver Operating Characteristic curve (AROC) values.

Results: The analysis identified 133 out of an initial set of 25 683 clinical diagnoses (represented by SNOMED codes) that were predictive of kidney cancer. The model achieved a cross-validated McFadden's R2 of 0.195 and an AROC of 0.799. Most of the identified codes are consistent with the known risk factors for kidney cancer.

Conclusions: It is possible to accurately predict the risk of kidney cancer from medical history using this method. Additional studies to establish high-dimensional predictive risk factors are needed to see if EHR personalized risk prediction can lead to cost-effective cancer screening and eventually better clinical outcomes.