Pelvic lymph node risk stratification in prostate cancer: A machine learning approach toward a predictive nomogram.

Abstract

349 Background: Pelvic lymph node involvement (PLNI) is a critical prognostic factor in PCa, guiding decisions regarding extended lymphadenectomy during radical prostatectomy (RP). Existing nomograms, such as Briganti and MSKCC, are widely used but face challenges, including limited specificity and variable accuracy depending on patient selection and practice heterogeneity. Machine learning (ML) approaches may enhance predictive accuracy and identify key features for future nomogram development. We aimed to develop a ML model for predicting the presence of PLNI in patients who underwent RP with pelvic lymphadenectomy (PL) at our institution. Methods: This was an observational retrospective study conducted at our institution’s PCa Clinical Care Center. We included patients who underwent RP with PL between 2015 and 2024. The primary endpoint was the presence of PLNI. Random Forest (RF) and Logistic Regression (LR) models were developed to predict PLNI. Model performance was evaluated through cross-validation using accuracy, precision, recall, F1-score, and area under the curve (AUC). Accuracy reflects overall correctness, precision indicates the proportion of true positives, recall measures sensitivity to positive cases, and the F1-score balances both metrics. The AUC summarizes overall discriminative ability across different classification thresholds. Feature importance was assessed using SHAP (SHapley Additive exPlanations) values, which quantify each variable’s contribution to model predictions. Results: After data curation, 982 patients and 19 preoperative clinical, biopsy, and imaging variables were included. The initial RF model achieved an accuracy of 0.85 but showed limited sensitivity for PLNI (recall 0.28), potentially translating into a high proportion of false negative cases. Removing less relevant D’Amico Risk Classification, recall improved to 0.62 with accuracy 0.80, F1-score 0.82 and AUC 74%. A final LR model using only PIRADS, PSA density, MRI lesion characteristics, and D’Amico total score achieved the best balance (accuracy 0.83, recall 0.67, F1-score 0.85 and AUC 80%). Feature importance consistently identified PIRADS, PSA density, Gleason biopsy, and number of positive cores as key predictors. Conclusions: This represents the first Latin American study to internally develop predictive models for PLNI in PCa. ML models demonstrated promising performance, with LR providing the most balanced combination of accuracy and sensitivity. Key predictive features identified may serve as the foundation for developing a clinically useful nomogram and online risk calculator. These findings demonstrate the feasibility of locally trained ML models to complement global nomograms and enhance personalized surgical decision-making in PCa. External validation and prospective evaluation are needed.