A seven-variable clinical prediction model for bone metastasis at initial diagnosis of prostate cancer

Decisions about bone-imaging workup at initial diagnosis of prostate cancer often have to be made before PSMA-PET/CT or other specialized investigations are available, particularly in resource-constrained settings. Many recent prediction models rely on advanced imaging, radiomics, specialized biomarkers, or large registry infrastructures and may therefore be difficult to apply in this earlier decision context. We retrospectively reviewed records of 291 consecutive men with newly diagnosed, histopathologically confirmed prostate adenocarcinoma admitted to the First Affiliated Hospital of Xinjiang Medical University between March 2011 and November 2023. Starting from 93 candidate predictors, we applied a five-stage hybrid selection procedure (univariable screening, exploratory LASSO, clinical prescreening, data-quality review, and confirmatory LASSO) to derive the final variable set. The retained predictors were entered into a multivariable logistic regression and visualized as a nomogram. Model behavior was characterized through discrimination, calibration, decision curve analysis, and 1,000-iteration bootstrap optimism correction. Of the 291 patients, 113 (38.8%) had bone metastasis. The final 7-variable model included anemia status, alanine aminotransferase, lactate dehydrogenase, alkaline phosphatase, age, Gleason score (three groups), and total prostate-specific antigen (tPSA). The apparent AUC was 0.736 (95% CI 0.674-0.799); the bootstrap bias-corrected AUC was 0.713 with a calibration slope of 0.933. This seven-variable model uses routinely available variables and showed moderate discrimination with stable calibration; external validation in independent samples is required before clinical use.