885 – Machine learning prediction of disease-free survival in breast cancer using blood cell counts and clinicopathological features: a multicentre, retrospective cohort study

Abstract

Introduction: Blood–based laboratory parameters, such as the neutrophil–lymphocyte ratio, have emerged as valuable
markers in several cancer types. Machine learning models have shown potential for developing predictive and prognostic biomarkers.
Objective: To assess the prognostic value of blood cell count and clinicopathological data when applied to machine learning models to predict disease-free survival in patients with breast cancer.
Methods: This multicenter, retrospective cohort study collected clinicopathological and complete blood count data from
2,864 patients diagnosed with breast cancer in six countries (Argentina, Brazil, Italy, Indonesia, Belgium, and Slovakia)
between January 1, 1999, and December 31, 2013. The database was split into training and validation sets at an 8:2 ratio.
A Cox proportional hazards model with LASSO regularization was used to select the prognostic features. Four survival
models were trained (RSF, XGB, GBM, and COX-LASSO) to estimate disease-free survival. This study was approved by the
Institutional Ethics Committee of the Federal University of São Paulo (decision number 7.205.965).
Results: Seven key characteristics were selected for model development: age, neutrophil–lymphocyte ratio, T3 stage,
positive lymph nodes, histological grade 3 (G3), estrogen receptor positivity, and human epidermal growth factor receptor 2 (HER2). Among the four machine learning models, the Gradient Boosting Machine (GBM) showed the best performance, achieving an area under the receiver operating characteristic curve (AUROC) of 0.610 (95% confidence interval
0.303–0.630). In the training cohort, the Random Survival Forest (RSF) model demonstrated the lowest overall Integrated
Brier Score (IBS=0.075).
Conclusion: A large and diverse cohort of women with breast cancer was analyzed, including clinical, pathologic, and
laboratory data. The final GBM model demonstrated consistent predictive performance across various cohorts. Further
prospective randomized controlled studies are needed to enhance the accuracy and clinical applicability of the model.
Ethics: Research Ethics Committee Approval (CAAE 80509824.9.1001.5505).
Registry: ClinicalTrials.gov (NCT06447532).