# Clinically Significant ISUP Upgrading in the Multiparametric MRI Era: Biopsy Tumor Burden Outperforms Complex Machine Learning Models in a Single-Center Exploratory Cohort

## TL;DR

This study finds that the amount of cancer in biopsy samples is a better predictor of aggressive prostate cancer than complex machine learning models.

## Contribution

A simple statistical model using biopsy tumor burden and PSA density outperforms complex ML models in predicting prostate cancer grade upgrading.

## Key findings

- Biopsy tumor burden (positive core ratio) was the strongest predictor of clinically significant upgrading.
- A simple logistic model outperformed machine learning classifiers in predicting cancer grade upgrading.
- Only 15.6% of patients experienced clinically significant upgrading in the study cohort.

## Abstract

Sometimes, a prostate biopsy underestimates how aggressive the cancer is. This study looked at ways to predict when the final surgery will find a higher-grade cancer than the initial biopsy. We analyzed 96 men who had both a biopsy and their prostate removed. We focused on factors like PSA (a blood test for prostate cancer), MRI scans, and biopsy results. We found that the extent of cancer involvement in biopsy cores, and to a lesser degree PSA density, were associated with upgrading risk to a more aggressive cancer at surgery. Using just these two factors, a simple statistical model predicted grade upgrading more accurately than more complex computer models. If confirmed in larger studies, this tool could help doctors identify patients who have more aggressive cancer than initially thought, so they can choose the best treatment.

Background/Objectives: Despite multiparametric MRI (mpMRI)-guided biopsy, clinically significant upgrading (CSU) of ISUP Grade Group (GG) at radical prostatectomy (RP) remains common in prostate cancer (PCa). We aimed to identify predictors of CSU (biopsy GG ≤ 2 to RP GG ≥ 3) using routine preoperative variables, and to benchmark a parsimonious logistic model against multiple machine learning (ML) classifiers. Methods: In this single-center exploratory analysis, 96 consecutive PCa patients underwent pre-biopsy mpMRI, systematic ± MRI-targeted biopsy, and RP. Predictive modeling was restricted to biopsy GG 1–2 patients (n = 64). LASSO-guided feature selection and Firth-penalized logistic regression were used to build a locked reference model, evaluated against ML classifiers using cross-validated discrimination, calibration, and decision curve analysis. Results: CSU occurred in 10/64 patients (15.6%). Positive core ratio was the dominant independent predictor (adjusted OR 1.54 per 10% increase, 95% CI 1.10–2.17). PSA density (PSAD) showed a consistent positive association but did not retain independent significance. The locked two-variable model (AUC ≈ 0.75–0.79) outperformed all ML classifiers in discrimination, calibration, and net clinical benefit; however, the limited event count (n = 10) constrains model stability, and these findings require external validation. Conclusions: In a PCa mpMRI-informed diagnostic pathway, CSU is primarily driven by biopsy tumor burden. A simple logistic model based on positive core ratio and PSAD outperformed more complex ML approaches in this exploratory cohort, supporting integration of biopsy tumor burden metrics into preoperative risk stratification pending external validation.