Artificial intelligence improves risk stratification for breast cancer recurrence and mortality in women exposed to pesticides: a call for reassessment of stratification criteria.

Introduction: Precision in clinical practice is essential for optimizing patient outcomes and quality of life. To enhance diagnostic accuracy and treatment efficacy, various healthcare studies - including those on breast cancer - have increasingly adopted machine learning (ML) techniques. By leveraging ML to analyze patient history data, researchers can predict disease outcomes more accurately and tailor treatments effectively. Brazil, the world's largest consumer of pesticides, faces significant public health challenges due to occupational exposure. Notably, pesticide exposure is not considered a risk factor in the current Diagnostic and Therapeutic Guidelines for Breast Carcinoma (Joint Ordinance No. 5, of April 18, 2019), which guides the diagnosis, treatment, and monitoring of breast cancer patients. In a recent study published by our group, we observed hidden risks associated with occupational pesticide exposure in women with breast cancer. The correlation between pesticide exposure and the severity of breast cancer in female farmers has already been demonstrated by our group previously. In this study, we focus on predicting the risk of death and cancer recurrence in these patients, comparing this population with patients diagnosed with cancer but not exposed to pesticides. Methods: In this context, the present study employed ML algorithms to predict the risk stratification for recurrence and mortality in breast cancer patients and to re-stratify them by incorporating pesticide exposure as an additional risk factor. Clinicopathological data from 427 women were used to train logistic regression, random forest, support vector machine, and gradient boosting, obtaining models to identify the algorithm with superior predictive performance. These models were applied to patient stratification, with pesticide exposure included as an additional parameter. Model performance was evaluated using precision, accuracy, recall, F1-score, and the area under the ROC curve (AUC-ROC). Results and discussion: Incorporating pesticide exposure data resulted in a 24.12% improvement in the prediction quality of the best model (random forest), demonstrating that ML models can better learn and understand patterns in the dataset when this risk factor is considered. These findings underscore the necessity of including pesticide exposure in risk stratification, particularly in regions of family farming.