Machine learning prediction of breast cancer local recurrence localization, and distant metastasis after local recurrences

Abstract

Abstract Local recurrences (LR) can occur within residual breast tissue, chest wall, skin, or newly formed scar tissue. Artificial intelligence (AI) technologies can extract a wide range of tumor features from large datasets helping in oncological decision-making. Recently, machine learning (ML) models have been developed to predict breast cancer recurrence or distant metastasis (DM). However, there is still a lack of models that consider the localization of LR as a tumor feature. To address this gap, here, we analysed data from 154 patients including pathological, clinical, and follow-up data (with an average follow-up of 133.16 months) on both primary tumors (PT) and recurrences. By using ML methods we predicted the localization of LR and the occurrence of DM after LR. The performance (ROC AUC) of the best ML models was 0.75, and 0.69 for predicting LR in breast parenchyma, and surgical scar tissue, respectively, and 0.74 for predicting DM after LR. We identified recurrence localization, and the time elapsed between the detection of primary breast carcinoma and the recurrence, and adjuvant chemotherapy as the most important features associated with further DM. We conclude that combining traditional prognostic factors with ML may provide important tools in the risk assessment of patients with breast LR.