Improving prostate cancer risk stratification using machine learning applied to routine PSA testing: a multicenter study

Background Prostate-specific antigen (PSA) screening reduces mortality but suffers from low specificity, leading to overdiagnosis and unnecessary biopsies. While specialized biomarkers exist, their high cost and limited availability hinder widespread implementation. Machine learning (ML) offers a scalable approach to enhance risk stratification using routinely collected clinical data. Methods We conducted a retrospective, large-scale multicenter study including 4,241 Brazilian men (709 prostate cancer cases; 3,532 controls) derived from a nationwide network spanning 421 clinical settings across 10 Brazilian states. Participants were men referred for prostate evaluation. Controls were defined by negative histopathology or PI-RADS 1–2 multiparametric MRI (diagnostic-evaluation cohort with a mixed reference standard). An interpretable machine learning model (PSA model) integrating total PSA, free-to-total PSA ratio, and age was evaluated on a held-out test set and, for robustness, by leave-one-group-out validation across four geographic groups. Incremental value from PSA density (PSAD), PSA kinetics, and complete blood count markers was also assessed. Results The PSA model significantly outperformed a fixed PSA threshold (≥4.0 ng/mL) on the test set (AUC of 0.73 vs. 0.63, p < 0.001). In the PSA gray zone (4–10 ng/mL), discrimination improved from 0.50 to 0.67. At the training-derived threshold, specificity increased from 0.44 to 0.60 (sensitivity 0.73 vs. 0.83), corresponding to 24% fewer referrals for MRI or biopsy at the cost of six additional missed ISUP ≥2 cancers, a potential benefit requiring prospective confirmation. Notably, the model showed higher detection of aggressive disease, with an AUC of 0.81 for intermediate-to-high-risk tumors (Gleason ≥7). Incorporating PSAD further enhanced performance (AUC 0.78; p=0.003), whereas PSA kinetics and CBC markers provided no significant incremental value. Across the four geographic groups, AUCs ranged from 0.70 to 0.74. Conclusions An interpretable ML model utilizing widely available clinical variables provides a scalable risk stratification tool for men referred for prostate evaluation. By demonstrating consistent discrimination across four geographic groups comprising 10 Brazilian states and hundreds of clinical settings, this framework offers a practical strategy to optimize resource allocation in real-world clinical practice.

Authors

Institutions

Publication Details

Journal
Frontiers in Oncology
Published
2026-11-19
DOI
https://doi.org/10.3389/fonc.2026.1994895
Primary Topic
Prostate Cancer Diagnosis and Treatment
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Improving prostate cancer risk stratification using machine learning applied to routine PSA testing: a multicenter study

Flávio Machado, Vinicius Moura Ribeiro, João V. M. Malvezzi, Pedro Henrique Araújo de Souza et al.
Frontiers in Oncology
Prostate Cancer Diagnosis and Treatment
article

Improving prostate cancer risk stratification using machine learning applied to routine PSA testing: a multicenter study

Flávio Machado, Vinicius Moura Ribeiro, João V. M. Malvezzi, Pedro Henrique Araújo de Souza, Marco Aurelio Kohara, Bruno Aragão Rocha, Daniella Castro Araújo, Maria Carolina Tostes Pintão, Otavio Jose Eulalio Pereira, Laura Rabelo Leite, Keity de Moura Pinho, Suzylaine da Silva Lima, Rejane Cristina dos Santos Silva, Felipe Vaz Peres, Flavia Helena da Silva
article en

Abstract

Background Prostate-specific antigen (PSA) screening reduces mortality but suffers from low specificity, leading to overdiagnosis and unnecessary biopsies. While specialized biomarkers exist, their high cost and limited availability hinder widespread implementation. Machine learning (ML) offers a scalable approach to enhance risk stratification using routinely collected clinical data. Methods We conducted a retrospective, large-scale multicenter study including 4,241 Brazilian men (709 prostate cancer cases; 3,532 controls) derived from a nationwide network spanning 421 clinical settings across 10 Brazilian states. Participants were men referred for prostate evaluation. Controls were defined by negative histopathology or PI-RADS 1–2 multiparametric MRI (diagnostic-evaluation cohort with a mixed reference standard). An interpretable machine learning model (PSA model) integrating total PSA, free-to-total PSA ratio, and age was evaluated on a held-out test set and, for robustness, by leave-one-group-out validation across four geographic groups. Incremental value from PSA density (PSAD), PSA kinetics, and complete blood count markers was also assessed. Results The PSA model significantly outperformed a fixed PSA threshold (≥4.0 ng/mL) on the test set (AUC of 0.73 vs. 0.63, p < 0.001). In the PSA gray zone (4–10 ng/mL), discrimination improved from 0.50 to 0.67. At the training-derived threshold, specificity increased from 0.44 to 0.60 (sensitivity 0.73 vs. 0.83), corresponding to 24% fewer referrals for MRI or biopsy at the cost of six additional missed ISUP ≥2 cancers, a potential benefit requiring prospective confirmation. Notably, the model showed higher detection of aggressive disease, with an AUC of 0.81 for intermediate-to-high-risk tumors (Gleason ≥7). Incorporating PSAD further enhanced performance (AUC 0.78; p=0.003), whereas PSA kinetics and CBC markers provided no significant incremental value. Across the four geographic groups, AUCs ranged from 0.70 to 0.74. Conclusions An interpretable ML model utilizing widely available clinical variables provides a scalable risk stratification tool for men referred for prostate evaluation. By demonstrating consistent discrimination across four geographic groups comprising 10 Brazilian states and hundreds of clinical settings, this framework offers a practical strategy to optimize resource allocation in real-world clinical practice.

Frontiers in OncologyVol. 16
Instituto Nacional de Câncer - INCA (BR), Fleury S.A. (Brazil) (BR)
Prostate Cancer Diagnosis and Treatment
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon