A newly developed machine-learning framework combining routine clinical markers demonstrates superior predictive performance in identifying prostate cancer among patients presenting with borderline prostate-specific antigen concentrations, according to findings from researchers at Dalian Medical University. The soft voting ensemble model, evaluated in a cohort of Chinese men with total serum prostate-specific antigen levels falling between 4 and 10 ng/mL, delivered an area under the receiver operating characteristic curve of 0.904 with a 95 percent confidence interval spanning 0.836 to 0.972. This diagnostic efficacy proved statistically significantly superior to that of any single standalone biomarker evaluated in the study.
Distinguishing between prostate cancer and benign prostatic hyperplasia has long represented a notable diagnostic challenge within this specific prostate-specific antigen concentration interval, frequently described in clinical urology as the diagnostic gray zone. Currently, clinicians lack reliable quantitative risk stratification tools capable of guiding individualized biopsy determinations for patients whose biomarker values sit within this intermediate spectrum.
All enrolled participants presented with baseline serum prostate-specific antigen measurements in the 4 to 10 ng/mL range.
These features comprised the free-to-total prostate-specific antigen ratio, the prostate health index, [-2]proPSA, prostate-specific antigen density, and total prostate volume.
The engineering of the machine-learning pipeline involved a soft voting ensemble algorithm configured through automated Optuna hyperparameter optimization. To counteract the pronounced numerical disparity between the 18 confirmed cancer cases and the 82 benign cases, the modeling framework incorporated specialized class imbalance correction techniques.
By maximizing Youden index, the investigators established an optimal decision cutoff value of 0.1726. At this diagnostic threshold, the ensemble model achieved an overall diagnostic accuracy of 81.00 percent, paired with a sensitivity of 94.44 percent and a specificity of 78.05 percent. Crucially for biopsy triage, the model demonstrated a negative predictive value of 98.46 percent, underscoring its ability to rule out malignancy with high reliability.
Subsequent evaluation using calibration curves and decision curve analysis further validated the practical utility of the machine-learning architecture. The calibration curve demonstrated close predictive consistency between estimated risk probabilities and observed pathological outcomes.
The investigators concluded that an ensemble model driven by five routine non-invasive clinical indicators offers dependable risk stratification for prostate cancer in Chinese men presenting with prostate-specific antigen levels in the 4 to 10 ng/mL range. Given its exceptionally high negative predictive value, the tool holds potential utility for accurately identifying low-risk individuals in whom invasive prostate biopsy could be safely deferred or avoided altogether.
Subject of Research: Machine learning ensemble risk prediction for prostate cancer biopsy decision-making in the PSA gray zone
Article Title: Clinical value of a soft voting ensemble model based on routine clinical indicators for risk prediction and biopsy decision-making of prostate cancer in Chinese men with PSA in the 4–10 ng/mL
Article References: Hua, Y., Li, Y., Xu, J., Wang, X., Li, X., & Wang, Y. (2026). Clinical value of a soft voting ensemble model based on routine clinical indicators for risk prediction and biopsy decision-making of prostate cancer in Chinese men with PSA in the 4–10 ng/mL. Frontiers in Oncology, 16, Article 1969579. https://doi.org/10.3389/fonc.2026.1969579
Image Credits: AI Generated
DOI: 10.3389/fonc.2026.1969579
Keywords: prostate cancer, prostate-specific antigen, PSA gray zone, machine learning, soft voting ensemble model, prostate biopsy, benign prostatic hyperplasia, prostate health index, free-to-total PSA ratio, p2PSA, PSA density, prostate volume, Optuna hyperparameter tuning, negative predictive value, decision curve analysis, repeated cross-validation, clinical decision support
News Source: Nathaniel Bowman. (November 20, 2026). Machine Learning Ensemble Accurately Predicts Prostate Cancer During Gray-Zone Biopsy Decisions. Scienmag.



