Scholarlink

Background <p>Clazosentan reduces angiographic vasospasm after aneurysmal subarachnoid hemorrhage (aSAH), but functional benefit may vary across patients. We used causal machine learning to explore heterogeneity and derive an interpretable rule.</p> Methods <p>In a secondary analysis of the RECOVER dataset [multicenter retrospective cohort of aSAH treated by clipping or coiling within 48&#xa0;h (<i>N</i> = 506)], we compared clazosentan-containing management (with or without fasudil) with fasudil-only prophylaxis. After applying inverse probability of treatment weighting for prespecified confounders [age, World Federation of Neurosurgical Societies (WFNS) grade, Fisher grade, and body mass index (BMI)], we used a causal forest to estimate conditional average treatment effects (CATEs) on favorable discharge outcome (modified Rankin Scale 0–2 at discharge). A policy tree summarized CATEs, and external validation was performed in an independent cohort (<i>N</i> = 181).</p> Results <p>CATEs were heterogeneous (mean 0.18 ± 0.14). The policy tree split first on BMI (≤ 20.03&#xa0;kg/m<sup>2</sup>): patients with BMI ≤ 20.03 and WFNS ≤ 2 showed no clear estimated benefit (mean CATE − 0.058), whereas those with BMI &gt; 20.03 or WFNS &gt; 2 showed higher estimated benefit (CATE 0.24–0.25). In external validation, the same rule identified a subgroup with higher odds of favorable recovery with clazosentan; estimates in the low-benefit subgroup were imprecise (<i>n</i> = 27).</p> Conclusions <p>In observational cohorts with limited overlap in treatment assignment, causal machine learning suggested heterogeneity in the estimated association of a clazosentan-containing strategy with discharge outcomes and produced a simple BMI/WFNS policy tree. These findings are hypothesis-generating and require prospective validation including safety endpoints and longer-term functional outcomes.</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

Heterogeneity of Clazosentan Benefit after Aneurysmal Subarachnoid Hemorrhage Demonstrated in an Externally Validated Causal Policy Tree

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月17日

查看全文
  • Kazuki Nishida,
  • Basile Chrétien,
  • Shinsuke Muraoka,
  • Takashi Izumi,
  • Masahiro Nishihori,
  • Shunsaku Goto,
  • Issei Takeuchi,
  • Ryosuke Tashiro,
  • Hiroyuki Sakata,
  • Denis Vivien,
  • Masaaki Mizuno,
  • Hidenori Endo,
  • Ryuta Saito

摘要

Background

Clazosentan reduces angiographic vasospasm after aneurysmal subarachnoid hemorrhage (aSAH), but functional benefit may vary across patients. We used causal machine learning to explore heterogeneity and derive an interpretable rule.

Methods

In a secondary analysis of the RECOVER dataset [multicenter retrospective cohort of aSAH treated by clipping or coiling within 48 h (N = 506)], we compared clazosentan-containing management (with or without fasudil) with fasudil-only prophylaxis. After applying inverse probability of treatment weighting for prespecified confounders [age, World Federation of Neurosurgical Societies (WFNS) grade, Fisher grade, and body mass index (BMI)], we used a causal forest to estimate conditional average treatment effects (CATEs) on favorable discharge outcome (modified Rankin Scale 0–2 at discharge). A policy tree summarized CATEs, and external validation was performed in an independent cohort (N = 181).

Results

CATEs were heterogeneous (mean 0.18 ± 0.14). The policy tree split first on BMI (≤ 20.03 kg/m2): patients with BMI ≤ 20.03 and WFNS ≤ 2 showed no clear estimated benefit (mean CATE − 0.058), whereas those with BMI > 20.03 or WFNS > 2 showed higher estimated benefit (CATE 0.24–0.25). In external validation, the same rule identified a subgroup with higher odds of favorable recovery with clazosentan; estimates in the low-benefit subgroup were imprecise (n = 27).

Conclusions

In observational cohorts with limited overlap in treatment assignment, causal machine learning suggested heterogeneity in the estimated association of a clazosentan-containing strategy with discharge outcomes and produced a simple BMI/WFNS policy tree. These findings are hypothesis-generating and require prospective validation including safety endpoints and longer-term functional outcomes.