Unsupervised Machine Learning for Five-Year Survival Prediction in Non-metastatic Breast Cancer Patients Treated with Chemotherapy, Radiotherapy, and Surgery

Abstract

Background: Breast cancer patients often receive a combination of surgery, chemotherapy, and radiotherapy; however, predicting post-treatment survival remains challenging. Although machine learning methods offer new opportunities for risk stratification, supervised models are constrained by incomplete data and predefined targets. By contrast, unsupervised approaches, such as association rule mining (ARM), can uncover hidden and clinically meaningful patterns in complex breast cancer datasets. Objectives: This study used ARM to identify patterns linking patient and tumor characteristics, as well as treatment details, to five-year survival or mortality in patients with non-metastatic breast cancer. Methods: We retrospectively analyzed clinical data from 198 patients with breast cancer. Continuous variables, including age, tumor size, and dose, were binned, and all variables were converted into transaction-style records. Apriori-based ARM was applied with thresholds of support ≥ 5%, confidence ≥ 60%, and lift > 1 to extract the top ten rules predicting survival status as “Dead” or “Alive.” Results: Association rule mining identified two high-risk profiles for five-year mortality: 1) HER2-positive invasive ductal carcinoma treated with standard-dose radiotherapy and/or chemotherapy and 2) medium-sized tumors (2 - 5 cm) with grade 2 pathology treated with mastectomy and chemotherapy. Two low-risk profiles with excellent five-year survival were also identified: 1) patients aged 50 - 69 years with small (< 2 cm), PR-positive tumors and negative margins and 2) patients aged 50 - 69 years with grade 3, HER2-negative, PR-positive tumors treated with mastectomy or standard-dose radiotherapy. Correlation analyses confirmed tumor size and chemotherapy as the strongest predictors of survival for ARM. Conclusions: Association rule mining identified distinct combinations of clinical and treatment factors that differentiated high-risk from low-risk breast cancer patients receiving combined therapies. These findings may help clinicians tailor follow-up intensity and treatment planning. Future studies should prospectively validate these rules and consider incorporating genomic or imaging data to further refine predictions.