Multi-cancer brain metastasis risk score development and validation using 220,246 whole transcriptomes and machine learning

2039 Background: Brain metastases occur in multiple cancer types with higher prevalence in lung, breast, melanoma, and GI cancers. The prognoses of patients who develop brain metastases are very poor and identification of brain metastasis risk could be useful for prognostication, monitoring, and therapy selection. Methods: Data from the whole transcriptome of 220,246 tumor profiles were analyzed and multiple machine learning models were trained on various molecular subtypes. The dataset was split 50% for training and the other 50% for testing, UMAP was employed for dimensionality reduction and the patterns learned across the entirety of the training dataset irrespective of brain metastasis were leveraged on the testing data set. Patients with brain metastasis were identified using the presence of ICD-10 code C79.31 (Secondary malignant neoplasm of the brain). As the absence of C79.31 could be due to the event not happening yet, patients without brain metastasis were stratified into groups based on 3, 4 or greater than 5 years without a C79.31 ICD-10 code. The brain metastasis risk score was defined by empirical evaluation of the positive predictive value in 7 groups of risk probabilities. The validation set contained 1,217 patients with brain metastasis and 4,631 without an observed brain metastasis within 3 years. Results: In the validation set, the prevalence of brain metastases within the risk scores across all cancer types ranged from 4% with the lowest risk score to 94% in the highest with 71% of cases receiving the lowest 2 risk scores, 15% the 2 intermediate risk scores, and 14% the 3 highest risk scores. For breast, lung and colon cancers, the prevalence of brain metastasis ranged from 4-10% in patients with the lowest risk scores to 92-100% in patients with the highest however the distribution of cases with each risk score was markedly different across cancer type. Breast cancer had 62% of cases receiving the lowest 2 risk scores versus 27% in lung, and 92% in colon. Breast cancer had 18% of cases receiving the 3 highest risk scores while lung had 42% and colon only 2% of cases with those 3 highest scores. Conclusions: Whole transcriptome data can be leveraged by a machine learning platform that employs dimensionality reduction techniques along with transfer learning to predict the risk of brain metastasis. This tool can be used to augment the clinical picture of cancer patients an unmet clinical opportunity to inform prognosis, monitoring, and therapeutic selection.