Machine Learning and Explainable Artificial Intelligence to Predict Occult Pelvic Nodal Metastases in Prostate Cancer.

Abstract

Determination of risk of occult pelvic lymph node involvement (LNI) in patients with cN0 prostate cancer is critical for determination of optimal treatment options. Though several nomograms exist, machine learning (ML) approaches might enable physicians to better assess individual risk by incorporating multiple clinical risk factors. Herein, we developed a ML model to predict occult LNI, and explained its composition using an explainable artificial intelligence (XAI) framework. Patients with cN0 prostate adenocarcinoma diagnosed from 2018-2020 were identified in the National Cancer Database. The query was limited to patients with known clinical staging and biopsy results who did not receive neoadjuvant therapy prior to pelvic nodal examination. Occult LNI was defined as pN1 disease based on surgical evaluation, with a minimum of 10 nodes examined. Five ML models were trained to predict LNI. Variables incorporated into the model were age, core biopsy results, Gleason scores, preoperative prostate specific antigen (PSA), and clinical T-stage. Model performance, measured using area under the receiver operator characteristic curve (AUC) on a holdout testing dataset, was compared to multivariable logistic regression. The best-performing model was explained using SHapley Additive exPlanation (SHAP) values. To permit more clinically-meaningful statistical interpretation, using a novel approach SHAP values were converted into odds ratios (OR), confidence intervals (CI), and p-values. A total of 23,131 patients met inclusion criteria; 2,676 (11.6%) had occult LNI. The Extreme Gradient Boosting model outperformed all other models with an AUC of 0.82 (95% CI: 0.78-0.86) compared to 0.80 (95% CI: 0.76-0.84) for logistic regression. Increasing PSA (OR: 1.031; p<0.001), number of positive biopsy cores (OR: 1.055; p<0.001), and percent positive biopsy cores (OR: 1.01; p<0.001) were all associated with increased risk of LNI. Based on observation of SHAP dependence plots, risk of LNI plateaued at PSA>20 ng/dL and >11 positive cores, while no plateau was observed for percent positive biopsy cores. Relative to T1c disease, patients with T3b were at highest risk of LNI (OR: 1.461; p = 0.003). Gleason score of 9 was associated with significant risk of LNI (Ref: Gleason 6; OR: 1.891; p<0.001). This was primarily driven by the primary Gleason score; primary Gleason 5 disease was associated with significant risk of LNI (Ref: Gleason 3; OR: 1.915; p<0.001) while a secondary Gleason score of 5 was the only grade with significant increased risk of LNI (Ref: Gleason 3; OR: 1.185; p = 0.004). Age and number of cores examined were not significant predictors of LNI. Our ML achieved improved performance relative to logistic regression at predicting occult LNI. XAI provided insight into the inner-working of the ML model. ML can be used to identify patients at risk for occult LNI and therefore inform clinical decision-making.