Brain metastasis is common among patients with lung cancer, occurring in up to 25%-50% of patients with metastatic lung cancer. Although there are multiple management approaches to brain metastasis, they vary in associated morbidity. Morbidity generally increases with the size and number of metastases. Increased frequency of brain MRI can enable early detection of brain metastasis at a smaller size but must be balanced against the costs and potential harms of overscreening. Our group developed a clinicogenomic brain metastasis prediction model, called Risk Assessment for Metastasis to Brain Outcome (RAMBO), to predict risk of future brain metastasis in patients with lung cancer to help guide MRI surveillance decisions.

The model was developed from patients with lung cancer seen at Stanford Medical Center. All patients had targeted panel sequencing of their lung cancer using Stanford's Solid Tumor Actionable Mutation Panel (STAMP) as part of their routine clinical care between January 2014 and June 2019. Patients were excluded if they did not have distant metastatic disease (either de novo stage IV or recurrence) and biopsy-proven lung cancer. Patients were further excluded if they had synchronous brain metastasis, defined as diagnosis of brain metastasis within 90 days of diagnosis with distant metastatic disease. The primary outcome was the time from date of diagnosis of distant metastatic disease to time of brain metastasis, death, or censoring.

We applied competing risk regression to obtain an estimate of the risk of brain metastasis. We considered a total of 45 features for inclusion in the model, including patient demographics (age, ethnicity, etc), clinical history (smoking history, etc), and tumor clinical and genomic characteristics (stage at diagnosis, histology, driver mutations, etc).  For feature selection, we applied machine learning approaches using the consensus from a set of penalized regression methods. On the basis of the selected features, we built the final model with complete-case analysis using a cause-specific proportional hazards model using death as competing risks. We used bootstrapping to obtain assessments of predictive accuracy.

The proposed model showed good calibration and high discrimination (bootstrapped AUC of 0.75; 95% CI, 0.64 to 0.84). When the study cohort was stratified into high-risk and low-risk groups using a 1-year risk threshold at the 85th risk percentile, the high-risk group had a significantly elevated observed incidence of developing brain metastasis versus the low-risk group (30.8% v 6.1% for 1-year incidence, P < .01). Compared with low-risk patients, patients at high risk of brain metastasis were younger and diagnosed at a more advanced stage and their cancer was more likely to have a central primary tumor location and have nonadenocarcinoma histology.

The goal of our model was to identify patients at high risk of brain metastasis who might benefit from a tailored intervention, such as increased frequency of brain MRI surveillance. Therefore, we further examined various clinical outcomes of the high-risk group identified by the proposed model stratified by MRI frequency. High-risk patients who had less frequent brain MRI showed larger brain metastasis compared with those who had brain MRIs more frequently  and were more likely to undergo surgery.