An unsupervised learning model that integrates clinical and MRI radiomics features outperforms existing models in predicting the 5-year progression-free survival of prostate cancer patients after prostatectomy: a multicenter study

Background Prostate cancer (PCa) is the second most common male cancer. Despite undergoing radical prostatectomy (RP), 20–30% of patients experience recurrence within 5 years. Unsupervised learning method based on radiomics features has proved its efficiency for predicting recurrence in patients with breast and lung cancer. In this study, we sought to identify subgroups of PCa patients after RP using an unsupervised clustering method based on clinical and MRI radiomics features, and further evaluate the prognostic value in predicting 5-year progression-free survival (PFS). Materials: Preoperative MRI and clinical data from 400 PCa patients (185 with recurrence) were collected from three centers (one training and two external validation groups). Radiomics features were extracted from index lesions. PFS-associated clinical and radiomics features were selected by least absolute shrinkage and selection operator (LASSO)-Cox analysis. The K-means clustering method was used to identify subgroups and construct a Radiomic-Clinical model. PFS differences across subgroups were assessed using Kaplan-Meier survival analyses. The predictive performance of the Radiomic-Clinical model was compared with the European Association of Urology (EAU), University of California, San Francisco (UCSF) Cancer of the Prostate Risk Assessment (CAPRA), and PIPEN models using the concordance index (C-index). Results Five clinical and 13 radiomics features were selected, and three distinct prognostic subgroups were identified within the Radiomic-Clinical model. The Radiomic-Clinical model demonstrated superior predictive accuracy with C-indices of 0.82 (training group), 0.78 (validation group 1), and 0.79 (validation group 2), outperforming the EAU (0.68, 0.70, 0.65), CAPRA (0.71, 0.67, 0.70), and PIPEN models (0.71, 0.70, 0.68) ( p  < 0.05). Conclusion Unsupervised learning using radiomics and clinical data effectively identifies distinct prognostic subgroups in PCa patients after RP, offering superior predictive performance over existing models for 5-year PFS.