Scholarlink

Purpose <p>Brain arteriovenous malformations (BAVM) pose a significant rupture risk, leading to morbidity and mortality. Identifying features associated with hemorrhagic presentation is crucial for improving understanding of AVM phenotypes at diagnosis. The purpose of this study is to comparatively evaluate logistic regression and machine learning models for identifying angioarchitectural features associated with hemorrhagic BAVM presentation, and to assess whether machine learning offers discrimination advantages over conventional regression.</p> Methods <p>We retrospectively analyzed 153 BAVM patients admitted to Stony Brook University Hospital between 2001 and 2022. A logistic regression model was built based on LASSO-selected variables. To compare analytical approaches, we evaluated several machine learning algorithms, and AdaBoost and SVM performed best and were selected for further analysis.</p> Results <p>Based on the multivariate regression analysis, BAVM size (OR = 0.95, 95% CI 0.91–0.99, <i>p</i> = 0.006), deep location (OR = 3.87, 95% CI 1.46–10.28, <i>p</i> = 0.007), presence of an arterial aneurysm (OR = 2.86, 95% CI 1.16–7.08, <i>p</i> = 0.023), and periventricular drainage (OR = 11.87, 95% CI 1.31–107.65, <i>p</i> = 0.028) were independent features associated with hemorrhagic presentation. Among machine learning models, AdaBoost achieved the highest discrimination (AUROC = 0.725), followed closely by SVM (AUROC = 0.719). Logistic regression achieved a comparable AUROC of 0.720, indicating that machine learning did not meaningfully improve discrimination over conventional regression in this cohort. SHAP analysis revealed that deep location, AVM size, and associated arterial aneurysms were consistently the most influential contributors across models.</p> Conclusion <p>Our study identified deep location, smaller AVM size, and arterial aneurysms are key features associated with hemorrhagic presentation across both regression and machine learning models. By examining both traditional statistics and explainable artificial intelligence, our study highlights the angioarchitectural features most consistently associated with hemorrhagic presentation across analytical approaches.</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

Angioarchitectural features of hemorrhagic presentation of brain arteriovenous malformations: A multivariate and explainable machine learning study

  • 期刊论文

  • 发表日期：

    2026年5月27日

查看全文
  • Seyed Morsal Mosallami Aghili,
  • Noah J. J. Lubin,
  • Alfonse J. Mazzarella,
  • Yindong Hua,
  • Jermaine Robertson,
  • Jade Basem,
  • Peter Gerace,
  • Jonathan Aminov,
  • Sujith Swarna,
  • Chiemeka Uwakwe,
  • Alan R Rheaume,
  • Michael Egnor,
  • Charles B. B. Mikell,
  • Sima Mofakham,
  • David J Fiorella

摘要

Purpose

Brain arteriovenous malformations (BAVM) pose a significant rupture risk, leading to morbidity and mortality. Identifying features associated with hemorrhagic presentation is crucial for improving understanding of AVM phenotypes at diagnosis. The purpose of this study is to comparatively evaluate logistic regression and machine learning models for identifying angioarchitectural features associated with hemorrhagic BAVM presentation, and to assess whether machine learning offers discrimination advantages over conventional regression.

Methods

We retrospectively analyzed 153 BAVM patients admitted to Stony Brook University Hospital between 2001 and 2022. A logistic regression model was built based on LASSO-selected variables. To compare analytical approaches, we evaluated several machine learning algorithms, and AdaBoost and SVM performed best and were selected for further analysis.

Results

Based on the multivariate regression analysis, BAVM size (OR = 0.95, 95% CI 0.91–0.99, p = 0.006), deep location (OR = 3.87, 95% CI 1.46–10.28, p = 0.007), presence of an arterial aneurysm (OR = 2.86, 95% CI 1.16–7.08, p = 0.023), and periventricular drainage (OR = 11.87, 95% CI 1.31–107.65, p = 0.028) were independent features associated with hemorrhagic presentation. Among machine learning models, AdaBoost achieved the highest discrimination (AUROC = 0.725), followed closely by SVM (AUROC = 0.719). Logistic regression achieved a comparable AUROC of 0.720, indicating that machine learning did not meaningfully improve discrimination over conventional regression in this cohort. SHAP analysis revealed that deep location, AVM size, and associated arterial aneurysms were consistently the most influential contributors across models.

Conclusion

Our study identified deep location, smaller AVM size, and arterial aneurysms are key features associated with hemorrhagic presentation across both regression and machine learning models. By examining both traditional statistics and explainable artificial intelligence, our study highlights the angioarchitectural features most consistently associated with hemorrhagic presentation across analytical approaches.