Machine learning–integrated routine biomarkers for breast cancer risk stratification: A scalable strategy for resource-limited settings.

Abstract

e12557 Background: Early breast cancer diagnosis is essential for improving survival outcomes and optimizing healthcare resource utilization. However, in resource-limited settings and low- and middle-income countries, access to screening programs and advanced risk assessment tools remains limited, and scalable strategies leveraging routinely collected data are underexplored. We evaluated routine blood-based biomarkers, individually and in combination, using a machine learning (ML) approach to develop a low-cost, scalable breast cancer risk assessment tool. Methods: This cross-sectional study included 236 women aged 40–70 years evaluated at Barretos Cancer Hospital between January 2024 and January 2025. Participants comprised newly diagnosed breast cancer patients (n = 125) and women undergoing screening with benign imaging findings (BI-RADS 1–2; n = 111). Routine laboratory markers included complete blood count, metabolic and inflammatory markers, thyroid and sex hormones, and tumor markers. Group comparisons were performed using Student’s t -test. A ridge regression model was trained using biomarkers significantly associated with breast cancer (p < 0.05). Results: The median age was 52 years. The majority of participants were White (51.7%), followed by Brown (37.1%) and Black (10.3%); 56.8% were postmenopausal. Several biomarkers were significantly higher in women with breast cancer (p < 0.05), including CA 15-3 (p = 0.018), CEA (p = 0.014), FSH (p = 0.016), T3 (p < 0.001) and Mean Corpuscular Hemoglobin Concentration - MCHC (p = 0.014). No strong correlations were observed among selected variables (maximum correlation 0.19 between CEA and CA 15-3). Incorporation of biomarkers with p < 0.05 (CA 15-3, CEA, FSH, T3 and MCHC) into the ridge regression model yielded a mean AUC of 0.72 (95% CI, 0.68–0.75) based on 1,000 iterations of stratified random subsampling, with sensitivity of 0.63 ± 0.08, specificity of 0.67 ± 0.08, accuracy of 0.65 ± 0.05, balanced accuracy of 0.65 ± 0.05, negative predictive value of 0.63 ± 0.05, and positive predictive value of 0.67 ± 0.06. Conclusions: An ML model combining routine blood-based biomarkers demonstrated potential as a low-cost pre-screening tool for breast cancer risk stratification. This approach may help prioritize diagnostic resources, particularly in resource-limited settings where access to advanced imaging and invasive diagnostic procedures is constrained. External validation in independent and diverse populations is required to confirm generalizability and support future clinical application.