Scholarlink

Background <p>Unruptured intracranial aneurysms (UIAs) are increasingly detected, yet individualized rupture risk prediction remains imprecise using conventional clinical scores. Machine learning (ML) models integrating high-dimensional imaging and clinical data have been proposed, but their incremental prognostic value remains uncertain. The objectives were to systematically evaluate ML-based models for predicting rupture or instability of UIAs and compare their performance with established clinical risk tools.</p> Methods <p>A systematic review and meta-analysis were conducted according to PRISMA 2020. PubMed, Embase, Scopus, Web of Science, and the Cochrane Library were searched (2010–2025). Studies developing or validating ML or deep learning models for rupture or growth prediction in UIAs were included. Discrimination was assessed using area under the receiver operating characteristic curve (AUC), and risk of bias using PROBAST-AI.</p> Results <p>Twenty-five studies comprising 18,569 aneurysms were included. ML models demonstrated good to excellent discrimination, with AUCs approximately 0.70–0.93. Performance increased with predictor complexity, with multimodal and imaging-rich models achieving the highest discrimination. In head-to-head comparisons, ML models consistently outperformed the PHASES score, whereas comparisons with logistic regression showed largely comparable performance without consistent ML superiority, particularly in externally validated cohorts. External validation was associated with modest attenuation of AUCs, and calibration reporting was infrequent.</p> Conclusions <p>ML-based models show discriminative capability for predicting rupture or instability in UIAs, driven primarily by enriched predictor domains rather than algorithmic complexity. However, limitations in external validation, calibration, and clinical utility reporting constrain translation into practice. Future research should prioritize prospective validation, calibrated absolute risk estimation, and integration into clinical decision-support frameworks.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Predicting rupture risk in unruptured intracranial aneurysms: a comparative evaluation of machine learning models and conventional risk scores

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月20日

查看全文
  • Donald E. Ogolo,
  • Enyereibe C. Ajare,
  • Brenda C. Opara,
  • Chinedu J. Onyemaobi,
  • Obioma Richard Akwada,
  • Kelechi Ndukuba,
  • Samuel C. Ohaegbulam

摘要

Background

Unruptured intracranial aneurysms (UIAs) are increasingly detected, yet individualized rupture risk prediction remains imprecise using conventional clinical scores. Machine learning (ML) models integrating high-dimensional imaging and clinical data have been proposed, but their incremental prognostic value remains uncertain. The objectives were to systematically evaluate ML-based models for predicting rupture or instability of UIAs and compare their performance with established clinical risk tools.

Methods

A systematic review and meta-analysis were conducted according to PRISMA 2020. PubMed, Embase, Scopus, Web of Science, and the Cochrane Library were searched (2010–2025). Studies developing or validating ML or deep learning models for rupture or growth prediction in UIAs were included. Discrimination was assessed using area under the receiver operating characteristic curve (AUC), and risk of bias using PROBAST-AI.

Results

Twenty-five studies comprising 18,569 aneurysms were included. ML models demonstrated good to excellent discrimination, with AUCs approximately 0.70–0.93. Performance increased with predictor complexity, with multimodal and imaging-rich models achieving the highest discrimination. In head-to-head comparisons, ML models consistently outperformed the PHASES score, whereas comparisons with logistic regression showed largely comparable performance without consistent ML superiority, particularly in externally validated cohorts. External validation was associated with modest attenuation of AUCs, and calibration reporting was infrequent.

Conclusions

ML-based models show discriminative capability for predicting rupture or instability in UIAs, driven primarily by enriched predictor domains rather than algorithmic complexity. However, limitations in external validation, calibration, and clinical utility reporting constrain translation into practice. Future research should prioritize prospective validation, calibrated absolute risk estimation, and integration into clinical decision-support frameworks.