Impact of surgical approach and survival prediction of malignant phyllode tumor by machine learning.

We aimed to analyze the effect of surgical approach on patients with malignant phyllode tumor of the breast (MPTB) and to develop a prognostic prediction model for patients with MPTB. We extracted MPTB patients aged 18-80 years between 2000 and 2020 from the SEER database. Covariable imbalance was reduced using the propensity-score matching (PSM) method. An analysis of Cox proportional hazard regression was performed to compare breast cancer-specific survival (BCSS) with overall survival (OS). The survival curves were generated using the Kaplan-Meier method. The 5-year BCSS and 5-year OS of patients with MPTB were predicted by ten models based on machine learning. According to multivariate Cox analysis, surgical treatment of MPTB does not affect long-term survival outcomes (p > 0.05). Among our study, the survival outcomes of mastectomy and BCS would not be statistically significant even for patients with poor pathologic type of MPTB (p > 0.05). In terms of AUC, CatBoost performed better than other algorithms with a 5-year BCSS of 0.8488 and a 5-year OS of 0.8512. BCS and mastectomy do not make a significant difference in the long-term survival outcomes of patients with MPTB. Therefore, we suggest that BCS is feasible and preferred provided that surgical margin requirements can be met. As a trusted model, CatBoost provides better guidance and support for the systemic treatment of patients with MPTB.