Predicting and Identifying Correlates of Inequalities in Breast Cancer Screening Uptake using National Data from India

Aleena Tanveer¹Raja Hashim Ali²Jitendra Majhi³Moumita Mukherjee^1*

• ¹Charité – Universitätsmedizin Berlin, Corporate Member of Freie Universität Berlin, and Humboldt-Universität zu Berlin, Berlin, Germany
• ²University of Europe for Applied Sciences, Potsdam, Germany
• ³All India Institute of Medical Sciences Kalyani, Kalyani, India

Despite large-scale national screening initiatives, breast cancer screening coverage in India remains extremely low; late-stage diagnosis contributes to high mortality among women. Creating accurate actionable predictions of socioeconomic and structural inequities in screening uptake is critical for designing equitable control strategies. This study applies machine learning to predict key determinants of screening uptake, and inequality using concentration indices and decomposition across economic, educational, and caste gradients. Using cross-sectional NFHS-5 data covering 68,526 women aged 30–49 years, variable selection was directed by Levesque’s healthcare access framework, spanning approachability, acceptability, affordability, availability, appropriateness. Three single learners—Logistic Regression, Naïve Bayes, Decision Tree—two ensemble learners—Random Forest, XGBoost—were trained on balanced-weighted data. To minimize overfitting risks following Synthetic Minority Oversampling Technique, model performance was validated through 10-fold cross-validation. Five evaluation metrics were compared to select the best-performing learner for predicting screening uptake. SHAP-based decomposition computed each feature-level contributions to inequalities across income, education, and social gradients. Screening uptake in India is exceptionally low (0.9%), with pronounced disparities across economic, educational, and social groups. Random Forest and XGBoost achieved high predictive accuracy (96%) and strong explainability (AUROC = 0.99), while Decision Tree yielded robust generalizability (mean AUROC = 0.995) in cross-validation splits. Feature importance analyses identified education, women’s autonomy, interactions with community health workers, spatial–provincial characteristics as major contributors to observed variation. Barriers such as distance to facilities, transport constraints, hesitancy in unaccompanied care seeking, and financial limitations showed smaller contributions to variance but remained relevant in shaping uptake patterns. Concentration indices confirmed pro-rich (0.1, p < 0.001), pro-educated (0.182, p < 0.001), and pro-marginalized (–0.011, p < 0.05) gradients. Tree-based decomposition suggested that affordability and education exacerbate pro-rich and pro-educated inequalities, yet could be an effective policy instrument to mitigate caste-based disparities. Access-related barriers intensified inequality across all gradients, whereas enabling access factors flattened them. Machine learning thus offers enhanced precision in predicting inequities in breast cancer screening uptake, revealing critical nonlinearities in gradients with access barriers. Findings highlight the need for financial protection, improved spatial accessibility, higher autonomy, strengthened community health worker engagement, and targeted awareness programmes for poor, less educated, socially marginalized women.