基于人工智能工具在前列腺癌诊断与管理中的系统综述与荟萃分析
该系统综述纳入2020年1月至2025年4月发表的43项研究,多数为回顾性诊断研究,深度学习是最常见模型(49%)。对34项研究的荟萃分析显示,诊断工具合并AUC为0.845(95%CI 0.809–0.881),预后工具为0.785(95%CI 0.715–0.856);深度学习模型优于经典模型(0.854 vs 0.805)。质量评估显示偏倚风险低,但工具适用性因验证方法仍存担忧。
要点
- AI诊断工具合并AUC 0.845,预后工具0.785,深度学习优于经典模型。
- 多数研究为回顾性、诊断导向,深度学习占49%。
- 偏倚风险低,但真实世界适用性因验证方法受限。
核心结论
- AI工具在前列腺癌诊疗全流程中表现良好,但汇总性能与真实世界适用性仍需关注。
研究设计与样本
- 系统综述与荟萃分析,检索PubMed、Scopus和ClinicalTrials.gov,纳入2020年1月至2025年4月发表的43项研究。
- 采用PRISMA 2020指南、Covidence软件和QUADAS-AI工具进行数据合成与质量评估,荟萃分析用R完成。
主要发现
- 43项研究中多数为回顾性诊断研究(n=29),深度学习是最常见AI模型(49%)。
- 34项研究随机效应荟萃分析:诊断工具合并AUC为0.845(95%CI 0.809–0.881,k=27)。
- 预后工具合并AUC为0.785(95%CI 0.715–0.856,k=7)。
- 亚组分析:深度学习模型AUC 0.854(95%CI 0.808–0.901,k=17)优于经典模型(XGBoost、SVM、RF)AUC 0.805(95%CI 0.756–0.856,k=14)。
- 7项叙述性研究提示大语言模型(LLM)的新兴作用;质量评估显示偏倚风险低,但验证方法引发适用性担忧。
关键数值与推荐
- 诊断工具合并AUC 0.845(95%CI 0.809–0.881)。
- 预后工具合并AUC 0.785(95%CI 0.715–0.856)。
- 深度学习模型AUC 0.854(95%CI 0.808–0.901),经典模型AUC 0.805(95%CI 0.756–0.856)。
临床怎么用
- AI工具可用于前列腺癌诊断与预后评估,深度学习模型性能更优,但需关注验证方法与真实世界适用性。
局限性
- 纳入研究多为回顾性设计,汇总性能与真实世界适用性存在担忧,验证方法影响工具适用性。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
BACKGROUND: Recent advancements in artificial intelligence (AI) hold great promise in oncology, including prostate cancer care. Despite its promises, there is a lack of comprehensive synthesis and knowledge regarding the efficacy of the current AI-based prostate cancer tools. This study aims to identify, evaluate and synthesize the existing evidence on AI-based tools developed for the diagnosis, prognosis, and management of prostate cancer. METHOD: We performed a systematic review of published studies from January 2020 to April 2025 that were retrieved from PubMed, Scopus, and Clinical Trials.gov focusing on the AI-based tools that are used in the diagnosis and management of prostate cancer care. Two independent reviewers utilized the PRISMA 2020 guidelines, develop a data charter and synthesize the study data using Covidence Software along with QUADAS-AI tool to assess paper quality and evaluate risk of bias. Meta-analysis was conducted on synthesized data using R. RESULTS: 43 studies were included, mostly retrospective and diagnostic-focused (n = 29), with deep learning being the most common AI model (49%). A meta-analysis of 34 studies with random effects pooled performance on AUC for the diagnostic tools (k = 27, MD = 0.845, 95% CI: 0.809,0.881), while prognostic tools (k = 7, MD = 0.785, 95% CI: 0.715, 0.856), with subgroup analysis indicating deep learning models (k = 17, MD = 0.854, 95% CI: 0.808, 0.901) out performed classical models (XGBoost, SVM, RF; k = 14, MD = 0.805, 95% CI: 0.756, 0.856). Seven narrative studies highlighted the emerging LLM role, and quality assessment revealed a low risk of bias, though concerns remained on the applicability of tools due to the validation method. CONCLUSION: This review highlights the promising AI tool performance for prostate cancer care continuum, while concerns on pool performances and real-world applicability. Future studies should emphasize human-centric design with equity-focused evaluations to ensure robust, ethical, scalable AI deployments in prostate cancer care.
引用
- Tun HM, Naing L, Malik OA, et al. Artificial Intelligence (AI)-based tools in the diagnosis and management of prostate cancer: a systematic review and meta-analysis. Prostate Cancer Prostatic Dis. 2025. PMID: 41388186. doi:10.1038/s41391-025-01060-w
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。