Unsupervised learning based on clinical factors and MRI radiomic features to predict 5-year progression-free survival in prostate cancer

Motivation: There is no recognized method in the world to accurately predict the risk of recurrence after radical prostatectomy. Goal(s): Find a new method to predict the risk of recurrence in prostate cancer patients. Approach: Preoperative bpMRI and clinicopathological information of 400 patients were collected from three centers. LASSO-cox analysis was used to select effective features. The k-means method was used to identify prognostic subgroups. K-M curves were plotted to compare the PFS of subgroups.The predictive efficacy of the model was assessed with concordance index. Results: Unsupervised learning can effectively identify high, medium, and low risk subgroups. Clinical-Radiomics model have higher predictive performance. Impact: Unsupervised learning-based bpMRI radiomics features and clinical factors have high predictive prognostic value, and these features have the potential to help to identify high-risk patients at an early stage, adjust the treatment regimen, and improve the prognosis of patients.