Using interpretable machine learning model to predict lymph node metastasis in patients with localized prostate cancer.

Background: This study aimed to develop and externally validate machine learning (ML) models for predicting lymph node metastasis (LNM) in patients with localized prostate cancer (PCa) using routinely available preoperative variables. Methods: Patients with localized PCa who underwent radical prostatectomy with extended pelvic lymph node dissection were retrospectively recruited from two institutions. The primary cohort comprised patients from the Affiliated Hospital of Qingdao University, while patients from the Third Affiliated Hospital of Soochow University were included as an external validation cohort. The primary cohort was randomly allocated into training (70%) and internal validation (30%) cohorts. Five ML algorithms were used to develop prediction models. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), and pairwise differences in AUC were assessed using DeLong's test. The optimal model was compared with the established Briganti and Memorial Sloan Kettering Cancer Center (MSKCC) nomograms. SHapley Additive exPlanations (SHAP) were utilized to enhance model interpretability and quantify the relative contribution of individual features. Results: A total of 632 patients were included in the primary cohort, and 395 patients were included in the external validation cohort. In the internal validation cohort, the LightGBM model significantly outperformed the other candidate ML algorithms (DeLong test, p < 0.05) and achieved an AUC of 0.8613 (95% CI: 0.7976-0.9251). In the external validation cohort, LightGBM maintained the highest AUC, reaching 0.8148 (95% CI: 0.7477-0.8819). Furthermore, the LightGBM model showed predictive performance comparable to that of the established Briganti and MSKCC nomograms. SHAP analysis demonstrated that total prostate-specific antigen, body mass index, biopsy gleason grade group, neutrophil-to-lymphocyte ratio, systematic prostate biopsy positive rate, neutrophil percentage-to-albumin ratio, clinical T stage, and platelet-to-lymphocyte ratio were the key contributors to the model's predictive performance. Conclusions: The LightGBM model incorporating eight key variables achieved competitive performance and showed predictive ability comparable to that of established Briganti and MSKCC nomograms. This model may serve as a useful preoperative decision-support tool, although prospective validation in larger and more diverse populations is still required before routine clinical implementation.