The Screening Equity Problem Current lung cancer screening eligibility criteria (USPSTF guidelines: age 50-80, 20 pack-year smoking history, current or former smoker) were derived primarily from predominantly white study populations. These criteria may systematically under-screen Black, Hispanic, and other minority individuals who develop lung cancer at lower smoking exposures.
Introducing EEC-GIFT This study presents EEC-GIFT: an Easy Ensemble Classifier with Group and Intersectional Fairness and Threshold optimization. It combines a specialized ensemble learning approach for imbalanced data (EEC) with algorithmic fairness preprocessing and post-processing techniques (GIFT) to build a screening model that is simultaneously accurate and equitable.
PLCO Dataset The model was trained and validated on the Prostate, Lung, Colorectal, and Ovarian (PLCO) Cancer Screening Trial dataset - 67,612 participants from a large, well-characterized randomized trial. PLCO includes demographic diversity and long-term lung cancer incidence outcomes, making it ideal for fairness-aware model development.
Dual Goals: Performance and Fairness Unlike most ML models that optimize only for accuracy, EEC-GIFT explicitly optimizes two objectives: maximizing sensitivity for lung cancer detection AND minimizing disparities in screening rates across demographic groups. These goals can conflict, requiring explicit trade-off management.
Easy Ensemble Classifier (EEC) Lung cancer is rare even in high-risk screening populations, creating severe class imbalance (few positives, many negatives). The Easy Ensemble Classifier addresses this by training multiple sub-classifiers on balanced subsets created by random undersampling of the majority class, then ensembling their predictions. This improves sensitivity for rare positive events.
Pre-Processing Fairness: Reweighting GIFT's pre-processing component reweights training samples to reduce historical disparities in data representation. Under-represented demographic groups receive higher sample weights during training, ensuring the model learns equally from all groups rather than optimizing primarily for the majority.
Post-Processing Threshold Optimization Rather than using a single decision threshold for all patients, GIFT uses group-specific thresholds calibrated to equalize equal opportunity disparity (EOD) - the difference in true positive rates between demographic groups. Different groups receive different screening probability cutoffs that equalize their chance of correct identification.
Intersectional Fairness Analysis Beyond single-attribute group fairness (e.g., race alone), GIFT evaluates intersectional fairness - disparities experienced by individuals at the intersection of multiple protected attributes (e.g., Black women, low-income Hispanic men). Intersectional analysis catches disparities invisible to single-attribute evaluation.
Sensitivity Improvement over USPSTF EEC-GIFT achieved an overall sensitivity of 85.16% for lung cancer detection, compared to 78.08% for the current USPSTF criteria. This 7 percentage point improvement means EEC-GIFT would identify substantially more lung cancers in a screening population without increasing the total number of people screened.
Equal Opportunity Disparity (EOD) The primary fairness metric - Equal Opportunity Disparity across racial and demographic groups - was 0.0034 for EEC-GIFT (p = 0.07, not statistically significant from zero, meaning unbiased). The USPSTF criteria showed an EOD of 0.0673 - nearly 20 times higher - indicating systematic underscreening of specific demographic groups.
Intersectional Fairness Maintained At the intersectional level (combinations of race, gender, socioeconomic status), EEC-GIFT maintained equitable screening rates across complex demographic subgroups where USPSTF showed meaningful disparities. This is a particularly important finding for health equity.
False Positive Rate Management Improving sensitivity must be balanced against false positives, which lead to unnecessary follow-up imaging with radiation exposure and patient anxiety. EEC-GIFT achieved its sensitivity gain while maintaining false positive rates at clinically acceptable levels, demonstrating that the fairness-performance trade-off can be managed.
Addressing Lung Cancer Disparities Black Americans develop lung cancer at lower cumulative smoking exposures than white Americans and are systematically under-screened by current pack-year-based criteria. EEC-GIFT's ability to incorporate diverse predictors (not just smoking history) enables more equitable risk assessment across populations with different smoking patterns.
Moving Beyond Pack-Year Criteria The model uses a broader set of features beyond age and pack-years, potentially including socioeconomic factors, occupational exposures, family history, and comorbidities. This multifactorial approach is more consistent with lung cancer's complex etiology than a simple smoking history threshold.
Policy Implications for Screening Guidelines If validated, EEC-GIFT type models could inform revision of national screening guidelines to incorporate ML-based risk assessment that explicitly accounts for equity. Several lung cancer risk prediction models already exist (PLCOm2012, BACH Pan-Canadian model), but none explicitly optimize for algorithmic fairness.
Fairness as a Model Requirement This study establishes an important precedent: that clinical ML models should be evaluated not just for accuracy but for disparate impact across demographic groups. Regulatory frameworks for clinical AI should require fairness auditing as a standard component of model evaluation.
PLCO Trial Limitations PLCO was conducted in the 1990s-2000s and reflects historical smoking patterns, cancer incidence rates, and screening modalities. Contemporary cohorts with lower-dose CT screening data and current smoking demographics may show different model performance.
Implementation Barriers Deploying group-specific thresholds raises potential ethical and regulatory concerns: using race or demographic characteristics as direct model inputs or decision factors can trigger legal and ethical scrutiny in healthcare systems. Implementation frameworks must address how demographic features are used and disclosed.
Access to Electronic Health Records The model requires comprehensive patient data (demographics, smoking history, comorbidities) that may not be uniformly available in all healthcare settings. Disparities in electronic health record completeness could inadvertently recreate access inequities.
Future Research Priorities The model requires external validation on contemporary screening cohorts including NLST and ongoing US/European lung cancer screening programs. Research should also explore whether the fairness gains persist when the model is deployed in real clinical environments with imperfect data, and how to communicate risk scores and screening recommendations equitably to patients from diverse backgrounds.