Rapid Fire Oral Presentation Clinical Oncology Society of Australia Annual Scientific Meeting 2026

A non-PSA risk prediction model for clinically significant prostate cancer: development and validation in a large Australian cohort (146534)

Pavla Vaneckova 1 , Qingwei Luo 1 , Michael David 1 2 , David P Smith 1 , Karen Chiam 1 , Rani Chand 1 , Chifeng Tan 1 , Visalini Nair-Shalliker 1
  1. The Daffodil Centre, The University of Sydney, and Cancer Council NSW, Woolloomooloo, NSW, Australia
  2. School of Medicine & Dentistry, Griffith University, Gold Coast, QLD, Australia

Prostate-specific antigen (PSA) testing is widely used for early detection of prostate cancer (PCa), but concerns about overdiagnosis and overtreatment limit its role in population screening. Existing risk prediction tools are largely PSA-based and estimate the risk of any PCa rather than clinically significant (csPCa) disease. We developed and internally validated the first Australian prediction model for csPCa without PSA, using personal and lifestyle factors only, to support shared decision-making and more targeted testing.

Data were obtained from the 45 and Up Study, a cohort of 267,357 participants recruited in New South Wales in 2005-2009. The analysis included 116,734 PCa-free men at baseline. Their records were linked to the NSW Cancer Registry, Admitted Patient Data Collection, Medicare claims (provided by Services Australia), NSW Registry of Births, Deaths and Marriages, and pathology datasets to identify no PCa, low-risk PCa, and csPCa outcomes. Predictors included age, family history of PCa, body mass index, smoking status, region of birth, history of vasectomy, and registry-confirmed melanoma. * A stacked ensemble machine-learning model was developed using training (80%) and test (20%) datasets, with internal validation in the test dataset.

Among 116,734 PCa-free men at baseline, 6,050 were diagnosed with csPCa, 2,018 with low-risk PCa and 1,254 with unknown-risk PCa during 10-year of follow-up. In the test dataset, the model demonstrated modest discrimination (AUC 0.64, 95% CI 0.60-0.64), with close agreement between predicted and observed risk, and good overall performance (multi-class Brier score 0.07). At the selected risk threshold, the model identified 95% of men later diagnosed with csPCa. Decision curve analysis showed greater net benefit across clinically relevant risk thresholds than testing all or no men.

This tool may support personalised decision-making about PSA testing in primary care by identifying men at higher risk of csPCa using personal risk factors alone.

 

*We-thank-the-Centre-for-Health-Record-Linkage-(CHeReL)-for-the-provision-of-linked-data. Secure-data-access-was-provided-through-the-Sax-Institute's-Secure-Unified-Research-Environment-(SURE).