(P-432) Machine Learning Applied to SEER Data for Prediction of Breast Cancer Specific Survival

Background: There has been a paucity in existing literature on the performance of machine learning (ML) methods including neural network (NN) in the prediction of survival among breast cancer patients.

Objectives: To determine risk factors and evaluate the performance of ML methods in the prediction of breast cancer survival.

Methods: ML methods were developed and validated in a retrospective cohort of female breast cancer with HER2 enriched (HR-/HER2+) and Triple Negative (HR-/HER2-) in 2012 in the Surveillance, Epidemiology, and End Results Research (SEER) Data, 18 Registries, 2000-2017. Patient characteristics including demographics, household income, AJCC stage and grade were assessed for analysis. The cancer cohort was split into 70:30 for training and validation (5-fold) and 100 overall runs. Gradient Boosting Classifier (GBC), LASSO, RIDGE, Random Forest (RF) and NN models were developed to predict and determine factors associated with breast cancer-specific death during the 5-year follow-up period. Mean (SE) C-statistics, precision and recall were utilized to evaluate the performance of each method. The Cox PH models with propensity score (PS) matching, IPTW and multivariate adjustment were used to estimate hazards ratios (HR) for cancer subtypes and risk factors.

Results: Among 7,019 patients with breast cancer included in the study, 22% were 70+ years old, 63% were white, and 8.5% were Asian. Localized only, regional and distant sides/nodes involved were present in 62%, 33.5% and 4.3% of patients, respectively. AJCC stages IAB, IIAB, IIIABC and IV were characterized in 38.5%, 42.5%, 15.5% and 4% of patients, respectively. The mean follow-up time was 51 months, and 16.5% patients died of breast cancer during the follow-up. GBC, LASSO, RIDGE, RF and NN models performed relatively well against external validation datasets with average C-statistics (SE) of 0.80 (0.012), both 0.79 (0.012), 0.76 (0.013) and 0.77 (0.001), respectively. Risk factors associated with low breast cancer survival included being elderly, black/Hispanic race, greater AJCC stages and grades. Variables of importance identified by ML models were consistent with risk factors from the Cox PH model.
Compared to HER2 enriched, an increased mortality risk of HR=2.0 (95% CI: 1.7-2.37) among patients with Triple Negative breast cancer. Cox PH models with multivariate or IPTW adjustment also provided consistent results.

Conclusions: The relatively strong performance of ML demonstrated potential benefits in prognostic evaluation for breast cancer–specific survival. A higher mortality risk was observed among Triple negative breast cancer patients in comparison to HER2 enriched type.