AI Models for Predicting Bone Metastasis in Prostate Cancer: A Systematic Review

Abstract

Background: Bone is the dominant hematogenous metastatic site in prostate cancer, and identifying patients at risk before or at staging could support more selective use of advanced imaging and earlier treatment planning. Artificial intelligence (AI), machine learning (ML), radiomics, and deep-learning approaches have increasingly been used to estimate bone-metastasis status from clinical variables and prostate imaging, but the evidence has not been synthesized around model validity and clinical readiness.

Methods: A systematic search was conducted to 31 July 2026 using PubMed/MEDLINE, scholarly web indexes, targeted publisher searches, and backward/forward citation checking. Eligible studies developed or validated an AI/ML/deep-learning or radiomics-based model for predicting the presence or risk of bone metastasis in adults with prostate cancer. Studies limited to detecting already visible bone lesions, prognosis after established bone metastasis, or inseparable non-bone metastatic outcomes were excluded. Reporting followed PRISMA 2020 principles. Model characteristics, predictors, validation strategy, and discrimination were narratively synthesized; methodological quality was appraised using PROBAST-informed domains.

Results: Fifteen studies published from 2019 to 2026 were included, representing 211,310 reported participant records, although 207,137 came from a single SEER-based study and several institutional cohorts may overlap. Twelve studies used MRI-derived radiomics, deep-learning features, or multimodal imaging; three were predominantly clinical ML models. Best reported validation/test discrimination ranged from 0.80 to 0.962. Four studies used an independent external cohort, with reported AUCs of 0.903-0.962. PSA, Gleason/ISUP grade, T and N stage, alkaline phosphatase, and radiomic signatures were recurrent predictors. Multimodal models frequently improved internal discrimination, but gains were less consistent on external data. Most studies were retrospective, geographically concentrated, and vulnerable to overfitting, class imbalance, non-independent cohorts, and inconsistent reference standards.

Conclusions: AI models show promising discrimination for prostate-cancer bone-metastasis risk, particularly when clinical variables are integrated with MRI-derived information. However, current evidence supports risk stratification and imaging triage rather than replacement of definitive staging. Prospective multicentre validation, standardized reference standards, calibration assessment, transparent reporting, and clinical-impact studies are required before routine deployment.