Prostate cancer is the second most frequently diagnosed cancer in men worldwide. A core challenge in its management is distinguishing clinically significant prostate cancer (csPCa) -- defined as ISUP grade group 2 or higher (Gleason score 7 or above) -- from indolent tumors that will never cause harm. Treating insignificant cancers unnecessarily exposes patients to serious side effects of surgery or radiation without benefit.
The PI-RADS v2.1 scoring system (Prostate Imaging -- Reporting and Data System) provides radiologists with a standardized 1-to-5 scale for assessing suspicious lesions on prostate MRI. PI-RADS has substantially improved risk stratification, but it has limitations: inter-reader variability is significant, particularly for intermediate PI-RADS 3 lesions, and the positive predictive value for csPCa even at PI-RADS 4 or 5 is far from perfect.
Prostate biopsy -- the definitive diagnostic test -- carries real risks including bleeding, infection, pain, urinary retention, and psychological distress. Studies have shown that multiparametric MRI (mpMRI) as a triage test before biopsy can detect csPCa with 93% sensitivity while reducing unnecessary biopsies by 27%. Yet combining MRI findings with clinical parameters in quantitative prediction models remains an area where machine learning can offer improvements over traditional logistic regression.
Most existing risk calculators -- including the Rotterdam and Barcelona MRI-based risk calculators -- rely on logistic regression. While these are interpretable and validated, they have a fundamental limitation: logistic regression captures only linear relationships between variables. Machine learning algorithms that model complex non-linear interactions and variable interdependencies may extract more predictive signal from the same data.
This retrospective study analyzed records from 4,799 men with serum PSA of 3.0 ng/mL or higher and/or abnormal digital rectal exam findings, recruited from 10 centers participating in the Catalan clinically significant prostate cancer early detection program between 2018 and 2022. All men underwent pre-biopsy multiparametric MRI followed by targeted and/or systematic biopsies.
Men receiving 5-alpha-reductase inhibitors (which artificially lower PSA) and those with prior prostate cancer, high-grade prostatic intraepithelial neoplasia, or significant comorbidities were excluded to reduce confounding. The outcome -- detection of csPCa (ISUP grade 2 or higher) at biopsy -- occurred in 41.1% of patients, a prevalence typical of a referral-based biopsy population.
Eight predictor variables were collected for each patient: age, PSA, prostate volume, PSA density (PSAD) (PSA divided by prostate volume measured on pre-biopsy MRI), digital rectal exam (DRE) findings, family history of prostate cancer, previous negative biopsy status, and PI-RADS score. This feature set is entirely composed of data routinely collected before a biopsy decision, requiring no additional testing.
The dataset was divided 80/20 into training (3,839 patients) and validation (960 patients) sets, with stratification to maintain equal proportions of csPCa cases in both sets. The 10 contributing centers showed substantial heterogeneity -- csPCa prevalence ranged from 35% to 65% across centers -- providing a realistic multi-institutional challenge for model development.
Eight classification models were developed and compared: logistic regression (stepwise variable selection), Ridge regression, LASSO (least absolute shrinkage and selection operator), elastic net, classification tree, random forest (300 trees), XGBoost (extreme gradient boosting), and a neural network (single hidden layer perceptron with 8 input and 8 hidden nodes). This spectrum spans classical statistical approaches, regularized linear models, tree-based ensembles, and neural networks.
Ridge, LASSO, and elastic net are regularization techniques that add penalty terms to the logistic regression objective function, forcing coefficients toward zero to reduce overfitting and handle multicollinearity. LASSO can zero out some coefficients entirely (performing automatic variable selection), while Ridge shrinks all coefficients proportionally. Elastic net combines both L1 (LASSO) and L2 (Ridge) penalties, with the best combination found at a mixing parameter of 0.4.
The XGBoost model was optimized with 800 trees, a learning rate of 0.2, maximum tree depth of 4, and an 80% subsample of training data and predictor variables per tree. The neural network used the hyperbolic tangent activation function and binary cross-entropy as its optimization objective. All models were trained only on training data and evaluated on the held-out validation set.
SHAP (SHapley Additive exPlanations) values were computed for all models to provide individual-patient-level interpretability. SHAP assigns each feature a contribution score for each prediction, derived from game theory principles that evaluate how adding each feature to all possible subsets of other features affects the model output. This transforms even complex ensemble models into clinically inspectable tools.
All models achieved good discrimination with AUC values above 0.8 on the validation set. The best-performing models were XGBoost and neural network (AUC 0.872 each), closely followed by logistic regression (AUC 0.870) and random forest (AUC 0.860). The differences between these top four models were not statistically significant, but all significantly outperformed the classification tree (AUC 0.807) and regularized linear models (Ridge, LASSO, elastic net: AUC 0.854 to 0.857).
When evaluated at a clinically relevant sensitivity of 0.90 (90% of csPCa cases correctly identified), XGBoost achieved the highest specificity of 0.640, meaning 64% of non-csPCa patients were correctly identified and could avoid biopsy. Random forest (0.638) and neural network (0.634) performed comparably, while logistic regression achieved 0.620. Importantly, none of the differences between the top models were statistically significant -- all performed similarly in terms of clinical utility.
Using the clinical utility curve, which plots the percentage of biopsies avoided against the acceptable rate of missed csPCa, the analysis showed that at a 10% misclassification threshold (accepting that 10% of csPCa cases fall below the cutoff), XGBoost could avoid 41.8% of unnecessary biopsies, random forest 41.7%, neural network 41.7%, and logistic regression 40.6%. These differences were not statistically significant between the top models.
Compared to a simpler clinical protocol -- performing biopsy for all patients with PI-RADS 3 or higher -- the machine learning models reduced biopsy rates by an additional 26 to 27 percentage points beyond what the PI-RADS threshold alone achieves. This represents a meaningful reduction in the burden of invasive testing while maintaining high sensitivity for significant cancer.
SHAP analysis consistently identified PI-RADS score as the most influential predictor across all models, particularly for patients with PI-RADS 4 or 5 lesions -- where MRI findings dominate the prediction and override most clinical variables. This aligns with clinical knowledge and validates that the models are learning from the most diagnostically meaningful signals.
For patients with intermediate PI-RADS 3 scores, where MRI alone is ambiguous, additional clinical variables become decisive. PSA density was the second most important feature overall, with the classification tree showing that for PI-RADS 4-5 patients with PSAD above 0.13, almost all qualified as csPCa. A suspicious DRE and positive family history of prostate cancer were highly influential for specific subgroups of patients -- individuals for whom these factors dramatically elevated risk even when other variables were intermediate.
A previous negative biopsy consistently emerged as a protective factor -- reducing the predicted probability of csPCa -- which is consistent with clinical evidence that repeat biopsies detect less cancer than first-time biopsies. SHAP enabled patient-level transparency: the example cases in the paper show how a 56-year-old with a PI-RADS 5 lesion and small prostate has csPCa predicted primarily by PI-RADS score and PSAD, while a 67-year-old with PI-RADS 3 and very large prostate (207mL) has non-csPCa predicted primarily by prostate volume driving low PSAD.
The SHAP analysis also demonstrated that the variable importance rankings were nearly identical across the four top models (random forest, XGBoost, logistic regression, neural network), suggesting that despite their algorithmic differences, all four models extracted the same underlying signal from the data. This consistency is reassuring for clinical adoption.
A key finding of this study is that despite the theoretical advantages of ensemble methods and neural networks, the performance gains over logistic regression were modest and not statistically significant. The AUC difference between the best models (XGBoost at 0.872) and logistic regression (0.870) was negligible. This suggests that with a well-structured dataset of 8 clinically curated variables, logistic regression captures most of the available signal.
The modest machine learning advantage becomes more apparent when examining clinical utility at high sensitivity. At 90% sensitivity, XGBoost's specificity (0.640) was meaningfully higher than logistic regression (0.620), suggesting that ensemble methods may better handle the complex interaction between PI-RADS, PSAD, and DRE at the extreme ends of the decision boundary -- the region that matters most clinically.
The classification tree model underperformed relative to ensemble methods (AUC 0.807), confirming a well-known finding in machine learning: individual trees overfit and lack the robustness of ensemble approaches. This study recommends that future clinical implementation prioritize random forest, XGBoost, or neural network models over single decision trees for prostate cancer risk stratification.
This study demonstrates that machine learning models combining PI-RADS scores and eight routine clinical variables can predict csPCa with AUC around 0.87 and reduce unnecessary biopsies by approximately 41-42% at a 10% false-negative tolerance. These models perform comparably to established tools like the Barcelona and Rotterdam risk calculators while offering enhanced interpretability through SHAP analysis.
The authors highlight that the emphasis on clinical utility -- the percentage of biopsies avoided at a given sensitivity -- is the most clinically meaningful metric, and one that many prior machine learning prostate cancer studies have neglected. Reporting AUC alone is insufficient; a model that achieves high sensitivity for csPCa at the cost of almost no specificity has no clinical utility, regardless of its AUC.
Key limitations include the retrospective design, the use of data from 10 heterogeneous centers without center-specific random effects adjustment, and the lack of external validation in a completely independent dataset outside Catalonia. Inter-reader variability in PI-RADS scoring remains a persistent challenge, and the model's performance may degrade in settings where MRI quality or reporting practices differ substantially from the training data. Future work should incorporate larger multi-national datasets and evaluate the model within existing clinical decision-making pathways.