Search

A critique of genomic risk classifiers is potential correlation with readily available clinical data. For these classifiers to enhance clinical decision making, they must demonstrate additional discrimination from clinical variables alone with regards to prognostic and/or predictive value. A 22-gene classifier is increasingly employed to inform treatment decisions in prostate cancer. This study aimed to assess whether its risk score remained independent of available clinical variables when using machine learning (ML) to predict genomic risk score outputs. This was a retrospective study of males with localized prostate cancer treated at one of twenty sites within a single hospital network. Patients whose tumors were sent for genomic risk profiling were eligible. Clinical features including year of biopsy, age, clinical stage, prostate specific antigen (PSA), Gleason score, and National Comprehensive Cancer Network (NCCN) risk group were extracted from the medical record and genomic risk score/category were extracted from the pathology results. Logistic regression for binary classification and linear regression for continuous classification plus 5 ML models were trained to predict the risk score, low-risk disease, and high-risk disease. Model performance was measured using area under the curve (AUC) for binary classification and Spearman rho (ρ) for regression. The best-performing model was explained using SHapley Additive exPlanation (SHAP) values. A total of 354 patients with biopsy specimens obtained between 2010 and 2024 were identified. Median age was 66.7 (IQR = 61.4-73.2). A total of 27.1%, 57.9%, and 15.0% of patients were NCCN low, intermediate, and high risk, respectively. Median genomic risk score was 0.385 (IQR = 0.26-0.58). A total of 57.6%, 18.1%, and 24.3% of patients had genomic risk classified as low, intermediate, and high, respectively. An extreme gradient boosting tree achieved the best performance at predicting genomic risk score (ρ: 0.526; 95% CI = 0.355-0.668). A random forest model achieved the best performance at predicting high-risk (AUC: 0.790; 95% CI = 0.671-0.909) and low-risk (AUC: 0.749; 95% CI = 0.631-0.867) genomic score. The most important variables for predicting risk score were primary Gleason, NCCN risk category, and total Gleason. Risk factors predicting high-risk disease included primary Gleason, NCCN risk group, and total Gleason. For low-risk disease they were primary Gleason, age, and total Gleason. ML predicted the output of genomic risk classifiers with favorable albeit imperfect performance using clinical variables alone. Future analyses should evaluate whether genomic risk classifiers may be particularly useful in the subset of patients whose genomic risk score differs from what was predicted using clinical variables alone. ML in combination with genomic risk should also be evaluated as synergistic tools to predict actuarial outcomes once sufficient follow-up is available. [ABSTRACT FROM AUTHOR]

Prediction of pathologic stage in patients with clinical node-negative (cN0) prostate cancer is critical for determination of optimal treatment options. Though several nomograms exist (with only the Partin tables capable of multi-class prediction [or classification within more than two categories]), machine learning (ML) approaches might allow for improved performance through their ability to capture and model complex interactions. Herein, we developed a multi-class ML model to predict pathologic stage and explained how variables were used to produce model predictions 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. Patients were classified as having organ confined disease (OC), extracapsular extension (ECE), seminal vesicle invasion (SVI), or lymph node involvement (LNI). These were considered mutually exclusive, so patients were classified as the highest stage they met criteria for. Variables incorporated into the model were age, core biopsy results, Gleason score, preoperative prostate-specific antigen (PSA), and clinical T-stage. Model performance was measured using one-versus-rest area under the receiver operator characteristic curve (AUC) on a holdout testing dataset. The best-performing model was explained using SHapley Additive exPlanation (SHAP) values. For model interpretation, SHAP values were converted into odds ratios (OR), confidence intervals (CI), and P values. A total of 23,131 patients met inclusion criteria; 8517 (49.5%), 4898 (28.4%), 1660 (9.6%), and 2140 (12.4%) had OC, ECE, SVI, and LNI, respectively. Random Forest model outperformed all other models with an AUC of 0.706 (95% CI = 0.691-0.720) compared to 0.691 logistic regression (95% CI = 0.676-0.705; p<.001) and the Partin Tables 0.674 (95% CI = 0.659-0.688; p<.001). The most important overall model features were primary Gleason, percentage of positive cores (PPC), and PSA. The most important features for each class are summarized in Table 1. Our ML achieved improved performance relative to logistic regression and the Partin Tables at predicting pathologic stage. XAI provided insight into the inner workings of the ML model. ML can be used to identify patients at risk for pathologic upstaging and therefore inform clinical decision-making. [ABSTRACT FROM AUTHOR]