Abstract: Objective  To identify responders with non-inferior efficacy of indobufen compared to aspirin in reducing the risk of stroke recurrence based on unsupervised learning method.
Methods  Based on the indobufen versus aspirin in acute ischemic stroke (INSURE) study, an unsupervised learning framework constructed using multiple kernel learning for dimensionality reduction and K-means clustering was developed to identify patients whose primary outcomes of indobufen were non-inferior to aspirin in treating acute moderate-to-severe ischemic stroke (i.e., responders). In the identified patients, the risks of secondary efficacy outcomes (new stroke within 1 year, new complex vascular events within 3 months and 1 year, new ischemic stroke within 3 months and 1 year, and adverse functional outcomes within 3 months and 1 year), safety outcomes (moderate-to-severe bleeding within 3 months), and secondary safety outcomes (moderate-to-severe bleeding within 1 year, any bleeding, and death within 3 months and 1 year) were evaluated for indobufen compared to aspirin. Differences of baseline data between responders and non-responders were compared.
Results  A total of 931 indobufen responders were identified (HR 0.70, 95%CI 0.46-1.06, Pnon-inferiority=0.004). Among the responders, the risks of secondary efficacy outcomes were lower in the indobufen group than in the aspirin group (new stroke within 1 year: HR 0.63, P=0.013; new ischemic stroke within 1 year: HR 0.68, P=0.039; new complex vascular events within 1 year: HR 0.64, P=0.014). The risk of any bleeding within 1 year was lower in the indobufen group than in the aspirin group (HR 0.41, P=0.025). The proportions of females, intracranial atherosclerotic stenosis, hypertension, diabetes mellitus, and dyslipidemia in the responders were significantly higher than in the non-responders. 
Conclusions  Unsupervised learning can be used to identify the responders with non-inferior efficacy of indobufen compared to aspirin in reducing the recurrence risk of  stroke. The proportions of females, intracranial atherosclerotic stenosis, and metabolic disorders in responders were higher than those in non-responders.

Key words: Unsupervised learning; Stroke; Ischemic stroke; Alternative antiplatelet therapy

摘要： 目的 基于无监督学习识别在降低卒中复发风险方面吲哚布芬疗效不劣于阿司匹林的应答患者。 
方法 基于吲哚布芬对比阿司匹林治疗急性缺血性卒中（indobufen versus aspirin in acute ischemic stroke，INSURE）研究，通过使用多核学习降维和K-means聚类构建的无监督学习框架，识别吲哚布芬在治疗急性中重度缺血性卒中时主要结局不劣于阿司匹林的患者（即应答患者）；在识别出的患者中，评估吲哚布芬相较于阿司匹林治疗发生次要疗效结局（1年内新发卒中、3个月和1年内新发复合血管事件、3个月和1年内新发缺血性卒中、3个月和1年内功能结局不良）、安全性结局（3个月内中重度出血）和次要安全性结局（1年内中重度出血，3个月和1年内任何出血、死亡）的风险；比较应答患者和非应答患者的基线资料差异。
结果 共识别出931例吲哚布芬应答患者（HR 0.70，95%CI 0.46～1.06，P非劣效=0.004）。在应答患者中，吲哚布芬组发生次要疗效结局的风险低于阿司匹林组（1年内新发卒中：HR 0.63，P=0.013；1年内新发缺血性卒中：HR 0.68，P=0.039；1年内新发复合血管事件：HR 0.64，P=0.014）；吲哚布芬组发生1年内任何出血的风险低于阿司匹林组（HR 0.41，P=0.025）。应答患者中的女性、颅内动脉粥样硬化性狭窄、高血压、糖尿病、血脂异常的比例显著高于非应答患者。
结论 无监督学习可以用于识别在降低卒中复发风险方面吲哚布芬疗效不劣于阿司匹林的应答患者。应答患者中女性、颅内动脉粥样硬化性狭窄和代谢性疾病的比例高于非应答患者。

关键词: 无监督学习; 卒中; 缺血性卒中; 替代抗血小板治疗