Understanding Cancer Risk Among Bangladeshi Women: An Explainable Machine Learning Approach to Socio-Reproductive Factors Using Tertiary Hospital Data

Breast cancer incidence is rising globally, and Bangladesh faces unique socio-economic and healthcare challenges, including late-stage diagnoses and limited screening. While reproductive factors such as age at menarche, parity, and contraceptive use are established risk determinants, their impact on hormone receptor-positive (HR+) and triple-negative breast cancer (TNBC) subtypes remains underexplored in low-resource settings. This study aims to identify key predictors of HR+ and TNBC subtypes among Bangladeshi women using machine learning models.

A case-control study was conducted at the National Institute of Cancer Research & Hospital (NICRH) in Bangladesh, involving 486 histopathologically confirmed breast cancer cases (246 HR+ and 240 TNBC) and 443 cancer-free controls. Socio-demographic and reproductive data were collected through structured interviews and medical records. Machine learning models, including Logistic Regression, SVM, Random Forest, and XGBoost, were used to predict breast cancer subtypes, with 5-fold cross-validation and Shapley values for evaluation and feature importance.

Among the cohort, rural residence was more common in both HR+ (76.4%) and TNBC (60.8%) cases compared to cancer-free individuals (53.7%). Lower education (≤5 years) was more prevalent in HR+ (80.5%) and TNBC (79.2%) cases than in the cancer-free group (20.1%) (OR = 6.42, 95% CI: 4.10–10.05, p < 0.001). TNBC patients had the highest rate of multiple abortions (≥2) (9.2%) (OR = 2.15, 95% CI: 1.30–3.55, p = 0.003) and postmenopausal status (59.2%) (OR = 1.88, 95% CI: 1.32–2.68, p = 0.001). Age at menarche was later in TNBC (13.03 years) compared to HR+ (12.05 years) and cancer-free women (12.64 years) (p < 0.001). XGBoost outperformed other models, achieving the highest sensitivity (0.750 for HR+ and TNBC) and F1 scores (0.750 and 0.706, respectively), with superior AUC performance. SHAP analysis identified key predictors, including cesarean delivery (HR+: OR = 1.92, 95% CI: 1.25–2.96, p = 0.002; TNBC: OR = 2.31, 95% CI: 1.46–3.65, p < 0.001), rural residence (HR+: OR = 2.65, 95% CI: 1.78–3.94, p < 0.001), and multiple abortions (TNBC: OR = 2.54, 95% CI: 1.48–4.37, p = 0.001). Feature analysis revealed that irregular menstruation and undernutrition were stronger predictors for HR+ cases, while wider gaps between reproductive milestones were more strongly associated with TNBC.

This study identifies socio-demographic and reproductive predictors for HR+ and TNBC subtypes among Bangladeshi women. Machine learning models offer valuable insights, emphasizing the need for tailored prevention strategies, improved healthcare access, and policy interventions.