光谱流式免疫图谱联合机器学习识别循环Th1样细胞作为晚期结直肠腺瘤血液标志物
该前瞻性研究纳入104例粪便隐血试验(FOBT)阳性接受结肠镜检查的患者,用光谱流式检测75种免疫细胞亚群的900个变量,结合五种监督机器学习模型分类。结果发现Th1样细胞是最主要预测变量,决策树模型对晚期息肉患者分类AUC达1.0(100%敏感性和特异性),全局模型准确率75%(p<0.01),宏观AUC 0.85。提示循环免疫分型可辅助FOBT优化结肠镜转诊。
要点
- Th1样细胞是区分晚期结直肠腺瘤的主要预测性免疫变量
- 决策树模型对晚期息肉分类AUC=1.0,敏感性和特异性均为100%
- 全局模型准确率75%(p<0.01),宏观AUC 0.85
核心结论
- 光谱流式免疫分型联合机器学习可从血样中识别携带晚期结直肠腺瘤的患者。
- Th1样细胞不仅是关键细胞,也是有前景的生物标志物,可补充FOBT以优先安排高危癌前病变患者结肠镜。
研究设计与样本
- 前瞻性招募104例FOBT阳性且接受结肠镜检查的患者。
- 按内镜和组织病理学分为无息肉组(NP,n=46)、非晚期息肉组(NA,n=33)和晚期息肉组(AP,n=25)。
- 用光谱流式表征75种免疫细胞亚群及其归巢、活化和耗竭谱,共产生900个变量。
- 用差异表达变量训练五种监督机器学习模型:随机森林、决策树、多项回归、多项式核支持向量机(SVM)和核K近邻。
主要发现
- 在91个差异表达免疫变量中,基于Boruta的特征选择识别出7个关键细胞群。
- Th1样细胞成为主导预测变量。
- 决策树模型总体性能最佳,对AP患者分类AUC=1.0,敏感性和特异性均为100%。
- 全局模型准确率达75%(p<0.01 vs 无信息率),宏观AUC为0.85。
关键数值与推荐
- 决策树模型对晚期息肉(AP)分类:AUC=1.0,敏感性100%,特异性100%。
- 全局模型准确率75%(p<0.01 vs 无信息率),宏观AUC 0.85。
临床怎么用
- 在FOBT阳性患者中,可考虑用光谱流式免疫分型联合机器学习识别晚期结直肠腺瘤高危者,以优化结肠镜转诊优先级。
- Th1样细胞可作为候选血液标志物,辅助补充现有FOBT筛查。
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: Colorectal cancer is the third most common malignancy worldwide. Although current screening relies on faecal occult blood testing (FOBT), its sensitivity for advanced adenomas (key precursor lesions) remains limited. Indeed, colonoscopy remains the gold standard for definitive diagnosis. Hence, development of novel blood-based screening tools in FOBT-positive patients is needed to optimize colonoscopy referral. METHODS: We prospectively recruited 104 FOBT-positive patients undergoing colonoscopy. Based on endoscopic and histopathological assessment, patients were classified into no polyps (NP, n = 46), non-advanced polyps (NA, n = 33) and advanced polyps (AP, n = 25). Seventy-five immune cell subsets and their homing, activation and exhaustion profiles were characterized by spectral cytometry, yielding 900 variables. Differentially expressed variables were used to train five supervised machine learning models including random forest, decision tree, multinomial regression, polynomial kernel support vector machine (SVM) and Kernel K-nearest neighbours. RESULTS: Among 91 differentially expressed immune variables, Boruta-based feature selection identified seven key cell populations. Th1-like cells emerged as the dominant predictive variable. The decision tree model achieved the best overall performance, with total classification of AP patients (AUC=1.0, 100% sensitivity and specificity). Global model accuracy reached 75% (p < 0.01 vs. no-information rate), with a macro-AUC of 0.85. CONCLUSIONS: This study demonstrates that spectral-cytometry immunotyping of the circulating immunome, combined with machine learning, can identify patients harbouring advanced colorectal adenomas from a blood sample. Hence, Th1-like cells emerge not just as key cells, but also as a promising biomarker that could complement current FOBT screening to prioritize colonoscopy in patients at highest risk of pre-malignant lesions.
引用
- Del Hierro AG, Izquierdo S, de Castro CG, et al. Spectral cytometry-based immune profiling coupled with machine learning identifies circulating Th1-like cells as a blood biomarker for advanced colorectal adenomas. Eur J Cancer. 2026. PMID: 42659836. doi:10.1016/j.ejca.2026.116980
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。