“A Population-Specific Breast Cancer Risk Prediction Model for Indian Women (A Pilot Study): Advancing Beyond Traditional Assessment Tools”

Background Breast cancer is the most prevalent cancer among women in India, characterized by late-stage diagnoses and high mortality rates. Existing breast cancer risk prediction models, such as the Gail and Tyrer-Cuzick models, were primarily developed using Western datasets, limiting their applicability to the Indian context due to socio-demographic, genetic, and cultural differences. Objective This pilot study aims to develop and validate a machine learning (ML)–based breast cancer risk prediction model tailored specifically to the Indian population, addressing the limitations of traditional tools, with the potential for future methodological expansion to build more robust and generalizable models. Methods A retrospective case-control pilot study was conducted using data from the National Cancer Institute (NCI)-AIIMS, comprising 590 breast cancer cases and 1,366 controls. Data preprocessing included cleaning, missing value imputation, and feature engineering of 66 clinical, genetic, and lifestyle factors. To address class imbalance and multivariate complexities, the XGBoost ensemble model was employed. Model performance was evaluated using accuracy, recall, precision, F1-score, and AUC-ROC metrics. Gini index values were used to interpret model predictions and identify key features for risk stratification. Results The model demonstrated robust predictive performance with an accuracy of 0.89 and AUC-ROC > 0.9, sensitivity of 73.95%, and specificity of 94.90% on the test dataset. Feature importance analysis enabled the development of a reduced model using the top 20 features, maintaining high accuracy and clinical relevance. The reduced model simplifies risk assessment in resource-limited settings. Conclusions This pilot study introduces a population-specific ML-based breast cancer risk prediction tool tailored to the Indian demographic. By incorporating culturally relevant variables and leveraging advanced machine learning techniques, the model addresses key limitations of Western-centric risk prediction tools. While the current methodology serves as an initial framework, it can be further expanded and refined to develop more robust and generalizable models for broader population coverage. Integration into clinical workflows and further validation across diverse Indian populations could transform early detection and personalized intervention strategies, significantly reducing the burden of breast cancer in India.