Is Systematic Biopsy Mandatory in All MRI-Guided Fusion Prostate Biopsies? A Machine Learning Prediction Model

TL;DR

A machine learning model can predict when systematic prostate biopsies are unnecessary, potentially reducing risks and costs while maintaining cancer detection rates.

Contribution

A novel machine learning model predicts when systematic biopsy is unnecessary in MRI-guided prostate biopsies.

Findings

The model achieved an area under the curve of 0.82 and a negative predictive value of 0.92.

Prostate volume, PSA density, age, and PSA were the most important predictors.

Omitting systematic biopsies based on the model would miss clinically significant cancer in 0.7% of cases.

Abstract

Prostate biopsies usually include targeted samples from magnetic resonance imaging lesions plus systematic samples across the prostate. Systematic sampling can detect extra cancers, but it adds needle cores and may increase bleeding, infection, discomfort, and costs, and often does not change treatment because decisions depend on the highest-risk cancer found. The authors analyzed 529 men who underwent both methods and built a machine-learning model from routine pre-biopsy information, including age, prostate size, prostate-specific antigen, and prostate-specific antigen density, to predict when systematic cores would contain a higher-risk cancer than targeted cores. In internal testing, the model showed good discrimination (area under the curve of 0.82) and a high negative predictive value (0.92); if systematic cores were omitted when the model predicted no benefit, clinically significant cancer would have been missed in 0.7% of patients. This approach could personalize biopsy schemes and reduce unnecessary cores, but it needs prospective validation.

Objectives: To develop a prediction model able to accurately predict which patients will harbor higher risk prostate cancer in the systematic biopsy template compared to the targeted biopsy during MRI/US fusion biopsy. Methods: We included patients who underwent fusion biopsy. Clinical and radiographic variables were collected from patients’ records. The outcome of the model was higher risk prostate cancer in the systematic compared with targeted biopsies. An extreme gradient boosting model was trained and tested. We evaluated variable importance and clinical benefit. Results: Five hundred and twenty-nine patients were included. Eighty-two (15.5%) patients had higher risk prostate cancer in the systematic biopsies. The area under the ROC curve and negative predictive value were 0.82 and 0.92, respectively. The four most important features for outcome prediction were prostate volume, PSAD, patient’s age, and PSA. The decision curve showed increased clinical benefit of our model at threshold probabilities of 0–0.5. Limitations include the retrospective design of the study and the lack of external validation of the model. Conclusions: We developed a prediction model able to accurately predict which patients must undergo systematic and targeted biopsy. This prediction model has the potential to help in the decision whether to perform SB and thus may lower the adverse event rate while keeping a high detection rate.

Figures