Predicting Risk of Malignant CNS Tumors From Medical History Events.

Abstract

Background and objectives: Malignant brain and other central nervous system tumors (MBT) are the second leading cause of cancer death among males aged 39 years and younger, and the leading cause of cancer death among males and females younger than 20. There are few widely accepted predictors and a lack of United States Preventive Services Taskforce recommendations for MBT. This study examined how medical history could be used to assess the risk of MBT.

Methods: Using over 400,000 patients' medical histories, including nearly 1,800 with MBT, Logistic Least Absolute Shrinkage and Selection Operator (LASSO) regression was used to predict MBT. More than 25,000 diagnoses were grouped into 16 body systems, plus pairwise and triple combinations, as well as indicators for missing values. Data were split into 80/20 training and validation sets with fit and accuracy assessed using McFadden's R2 and the area under the receiver operating characteristic curve (AUC).

Results: Diagnoses of the endocrine, nervous, and lymphatic systems consistently showed greater than three times more association with MBT. The best performing model at an AUC of 0.83 consisted of 14 body system diagnosis groups and pairwise interactions among groups, in addition to demographic, social determinant of health, death, and six missing diagnosis grouping indicators.

Conclusions: This study demonstrated how large data models can predict MBT in patients using EHR data. With the lack of preventive screening guidelines and known risk factors associated with MBT, predictive models provide a universal, non-invasive, and inexpensive method of identifying at-risk patients.