Evaluating machine learning model bias and racial disparities in non-small cell lung cancer using SEER registry data.

Background: Despite decades of pursuing health equity, racial and ethnic disparities persist in healthcare in America. For cancer specifically, one of the leading observed disparities is worse mortality among non-Hispanic Black patients compared to non-Hispanic White patients across the cancer care continuum. These real-world disparities are reflected in the data used to inform the decisions made to alleviate such inequities. Failing to account for inherently biased data underlying these observations could intensify racial cancer disparities and lead to misguided efforts that fail to appropriately address the real causes of health inequity.

Objective: Estimate the racial/ethnic bias of machine learning models in predicting two-year survival and surgery treatment recommendation for non-small cell lung cancer (NSCLC) patients.

Methods: A Cox survival model, and a LOGIT model as well as three other machine learning models for predicting surgery recommendation were trained using SEER data from NSCLC patients diagnosed from 2000-2018. Models were trained with a 70/30 train/test split (both including and excluding race/ethnicity) and evaluated using performance and fairness metrics. The effects of oversampling the training data were also evaluated.

Results: The survival models show disparate impact towards non-Hispanic Black patients regardless of whether race/ethnicity is used as a predictor. The models including race/ethnicity amplified the disparities observed in the data. The exclusion of race/ethnicity as a predictor in the survival and surgery recommendation models improved fairness metrics without degrading model performance. Stratified oversampling strategies reduced disparate impact while reducing the accuracy of the model.

Conclusion: NSCLC disparities are complex and multifaceted. Yet, even when accounting for age and stage at diagnosis, non-Hispanic Black patients with NSCLC are less often recommended to have surgery than non-Hispanic White patients. Machine learning models amplified the racial/ethnic disparities across the cancer care continuum (which are reflected in the data used to make model decisions). Excluding race/ethnicity lowered the bias of the models but did not affect disparate impact. Developing analytical strategies to improve fairness would in turn improve the utility of machine learning approaches analyzing population-based cancer data.