A Systematic Review of Artificial Intelligence‐Based Clinical Decision Support Systems in Prostate Cancer Management

Abstract

Background: Prostate cancer (PCa) affects over 1.4 million individuals annually and remains one of the leading causes of cancer-related death globally. Artificial intelligence-based clinical decision support systems (AI-CDSS) are increasingly used to improve diagnostic and treatment decisions in PCa management. Objective: This systematic review aims to evaluate the development, implementation, and limitations of AI-based CDSS applied to the diagnosing, staging, treatment, or management of PCa. Method: The review protocol was registered with PROSPERO (ID = CRD42024498335) and adhered to preferred reporting items for systematic reviews and meta-analyses (PRISMA) guidelines. Using the (population/intervention/comparison/outcome) PICO framework, research questions were formulated to examine system development, AI models applied, adherence to CDSS standards, and integration into clinical decision-making phases. A systematic search of PubMed, Scopus, Web of Science and Embase retrieved 1137 articles; after removing 173 duplicates, 964 records were screened. A subsequent update search in June 2024 identified no additional papers. Two reviewers independently screened titles and abstracts, full texts, and resolved disagreements through discussion with a third reviewer. Studies were included if they developed and evaluated AI-based CDSS for PCa with reported performance metrics. Studies were excluded if they did not involve AI-based CDSS, lacked evaluation data, or did not address clinical decision-making. Results: Ten studies met the inclusion criteria. Cohorts ranged from 10 to 1.93 million records. Reported tasks included diagnosis, risk/recurrence prediction, survival estimation, and radiotherapy/ brachytherapy plan optimization. Common models were gradient boosting, neural networks, and deep reinforcement learning (DRL). Performance was variable but modern ML generally outperformed traditional baselines for survival and risk prediction (e.g., AUROC ≈ 0.84 for three-year recurrence; C-index up to 0.92 for five-year progression-free survival (PFS)). A workflow CDSS without predictive modeling reduced decision-making time by 60%. Key limitations included small or heterogeneous samples, limited generalizability, inconsistent or incomplete data inputs, scarce external validation/calibration, and interpretability concerns. Core CDSS components were inconsistently implemented. Conclusions: AI-based CDSS for PCa show promise for diagnosis, prognosis, and treatment planning and may improve efficiency. However, broader adoption will require high-quality, multi-site data; external validation and calibration; interpretable models; and robust integration with EHRs and clinical workflows. Prospective clinical evaluations are needed to confirm patient-level impact.

Ähnliche Arbeiten