6D-02 - Developing a Cancer Stage Model in Patients with Incident Colorectal Cancer Using Data from a Cancer Care Quality Program and Administrative Claims

Background: Cancer stage is an important factor in research and clinical care because it impacts treatment choices and survival. However, administrative medical claims lack explicit codes on cancer stage.

Objectives: To evaluate statistical and machine learning models to predict stage IV (metastatic) colorectal cancer (CRC) using claims data.

Methods: The study population included U.S. adult commercial health plan members from the Healthcare Integrated Research Database® whose first diagnosis of CRC occurred between July 2016 and April 2023, and whose oncologist recorded cancer stage in a cancer care quality program (CCQP) database, representing the study’s "gold” standard. Seven models predicting stage IV (versus stage I-III) CRC were developed, one based on metastatic cancer diagnosis codes alone, and others that incorporated patient demographics, comorbidity scores, symptoms, diagnostic evaluations, other diagnoses, office and facility visits, and treatments, using various machine learning models. We developed the models in a training dataset (75% random sample) and evaluated the testing data based on sensitivity, specificity, and area under the receiver operating characteristic curve (AUC-ROC).

Results: The cohort included 6,408 patients with CRC, among whom 3,498 were in stages I-III and 2,910 in stage IV. The sensitivity, specificity, and AUC-ROC score for each model were as follows: logistic regression with metastatic cancer diagnosis only (0.85, 0.88, 0.87), parsimonious logistic regression (0.94, 0.90, 0.96), elastic net (0.93, 0.90, 0.96), random forest (0.94, 0.89, 0.96), random forest including Agency for Healthcare Research and Quality variables and drug subclass (0.91, 0.85, 0.94), gradient boosted (0.86, 0.93, 0.95), and SuperLearner ensemble (0.89, 0.95, 0.96), respectively.

Conclusions: Using administrative claims data, with CCQP oncologist-recorded cancer stage as the gold standard, diagnosis of metastatic cancer at distant sites was the strongest predictor of stage IV CRC (AUC-ROC: 0.87). Multivariable machine learning models enhanced the ability to distinguish between stage IV and other cancer stages, with factors such as colon resection, liver biopsy, immunotherapy, and radiation therapy improving the AUC-ROC score by 9 percentage points (AUC-ROC: 0.96); however, little difference in AUC-ROC scores was observed among the machine learning models. The logistic regression model was the most easily interpretable and performed as well as more complex models. This method can be applied to other cancer types to increase the value of claims data for cancer research and care.