Novel models based on machine learning to predict the prognosis of metaplastic breast cancer.

Abstract

Background: Metaplastic breast cancer (MBC) is a rare and highly aggressive histological subtype of breast cancer. There remains a significant lack of precise predictive models available for use in clinical practice.

Methods: This study utilized patient data from the SEER database (2010-2018) for data analysis. We utilized prognostic factors to develop a novel machine learning model (CatBoost) for predicting patient survival rates. Simultaneously, our hospital's cohort of MBC patients was utilized to validate our model. We compared the benefits of radiotherapy among the three groups of patients.

Results: The CatBoost model we developed exhibits high accuracy and correctness, making it the best-performing model for predicting survival outcomes in patients with MBC (1-year AUC = 0.833, 3-year AUC = 0.806; 5-year AUC = 0.810). Furthermore, the CatBoost model maintains strong performance in an external independent dataset, with AUC values of 0.937 for 1-year survival, 0.907 for 3-year survival, and 0.890 for 5-year survival, respectively. Radiotherapy is more suitable for patients undergoing breast-conserving surgery with M0 stage [group1: (OS:HR = 0.499, 95%CI 0.320-0.777 p < 0.001; BCSS: HR = 0.519, 95%CI 0.290-0.929 p = 0.008)] and those with T3-4/N2-3M0 stage undergoing mastectomy [group2: (OS:HR = 0.595, 95%CI 0.437-0.810 p < 0.001; BCSS: HR = 0.607, 95%CI 0.427-0.862 p = 0.003)], compared to patients with stage T1-2/N0-1M0 undergoing mastectomy [group3: (OS:HR = 1.090, 95%CI 0.673-1.750 p = 0.730; BCSS: HR = 1.909, 95%CI 1.036-3.515 p = 0.038)].

Conclusion: We developed three machine learning prognostic models to predict survival rates in patients with MBC. Radiotherapy is considered more appropriate for patients who have undergone breast-conserving surgery with M0 stage as well as those in stage T3-4/N2-3M0 undergoing mastectomy.