Machine learning prediction of pathological complete response and overall survival of breast cancer patients in an underserved inner-city population

Abstract

BackgroundGeneralizability of predictive models for pathological complete response (pCR) and overall survival (OS) in breast cancer patients requires diverse datasets. This study employed four machine learning models to predict pCR and OS up to 7.5 years using data from a diverse and underserved inner-city population.MethodsDemographics, staging, tumor subtypes, income, insurance status, and data from radiology reports were obtained from 475 breast cancer patients on neoadjuvant chemotherapy in an inner-city health system (01/01/2012 to 12/31/2021). Logistic regression, Neural Network, Random Forest, and Gradient Boosted Regression models were used to predict outcomes (pCR and OS) with fivefold cross validation.ResultspCR was not associated with age, race, ethnicity, tumor staging, Nottingham grade, income, and insurance status (p > 0.05). ER−/HER2+ showed the highest pCR rate, followed by triple negative, ER+/HER2+, and ER+/HER2− (all p < 0.05), tumor size (p < 0.003) and background parenchymal enhancement (BPE) (p < 0.01). Machine learning models ranked ER+/HER2−, ER−/HER2+, tumor size, and BPE as top predictors of pCR (AUC = 0.74–0.76). OS was associated with race, pCR status, tumor subtype, and insurance status (p < 0.05), but not ethnicity and incomes (p > 0.05). Machine learning models ranked tumor stage, pCR, nodal stage, and triple-negative subtype as top predictors of OS (AUC = 0.83–0.85). When grouping race and ethnicity by tumor subtypes, neither OS nor pCR were different due to race and ethnicity for each tumor subtype (p > 0.05).ConclusionTumor subtypes and imaging characteristics were top predictors of pCR in our inner-city population. Insurance status, race, tumor subtypes and pCR were associated with OS. Machine learning models accurately predicted pCR and OS.