文献速递小儿神经外科与脑积水神经重症

无创快速机器学习增强型颜色脑脊液诊断在脑室外引流感染筛查中的概念验证研究

Proof-of-concept study of noninvasive, rapid, machine learning-enhanced, color-based CSF diagnostics: a novel approach to external ventricular drain infection screening.
PubMed 文献 · J Neurosurg · 2025-09-19

subspace KNN 80-20分割验证准确率90.4%,随机森林留一患者交叉验证准确率81.1%。装置成本低于100欧元,高特异性或可减少不必要的脑脊液采样。

要点

  • BOSoMetre低成本连续监测EVD脑脊液颜色,机器学习区分感染与非感染样本。
  • subspace KNN 80-20验证AUC 0.968,随机森林LOOCV验证AUC 0.736。
  • 高特异性有望减少不必要的脑脊液采样及相关医源性感染风险。

核心结论

  • BOSoMetre作为低成本、开源工具在连续脑脊液监测和早期感染检测方面显示潜力,尤其适用于资源有限环境。

研究设计与样本

  • 作者设计BOSoMetre装置,使用微控制器和TCS3200颜色传感器置于3D打印腔室内,通过EVD管连续捕获红、绿、蓝和透明通道的实时光学测量。
  • 2024年10月至2025年1月前瞻性纳入20例因梗阻性脑积水或感染需放置EVD的患者,最终15例纳入分析。

主要发现

  • subspace KNN分类器80-20分割验证准确率90.4%,灵敏度92%,特异度90.4%,AUC 0.968。
  • 随机森林留一患者交叉验证准确率81.1%,灵敏度71.5%,特异度89.2%,AUC 0.736。
  • 装置成功区分清洁与感染脑脊液样本,识别非感染样本的特异性尤其高。

关键数值与推荐

  • 脑脊液样本按美国感染病学会2017年指南分类。
  • subspace KNN采用80-20分割验证,随机森林采用留一患者交叉验证(LOOCV)。

临床怎么用

  • 该装置成本低于100欧元且开源,适合资源有限环境连续监测脑脊液。
  • 高特异性可能减少不必要的脑脊液采样及关联的医源性感染风险。

局限性

  • 初步结果需在更大队列中进一步验证以确认临床效用。
  • 需克服本概念验证研究中发现的技术局限性。

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

OBJECTIVE: The objective was to develop and validate a proof-of-concept, low-cost, noninvasive device capable of continuously monitoring CSF in external ventricular drainage systems in order to enable earlier detection of infections. METHODS: The authors designed BOSoMetre (CSF-o-Meter), a device that uses a microcontroller and TCS3200 color sensor housed in a 3D-printed chamber for continuous CSF monitoring. The system captures real-time optical measurements across red, green, blue, and clear channels through the external ventricular drain (EVD) tube. Between October 2024 and January 2025, the authors prospectively enrolled 20 patients requiring EVD placement for obstructive hydrocephalus or infection, with 15 included in the final analysis. CSF samples were classified according to Infectious Diseases Society of America 2017 guidelines. The authors processed approximately 4.8 million sensor readings and applied machine learning algorithms using two validation approaches: the subspace k-nearest neighbors (KNN) classifier with 80-20 split validation, and random forest with leave-one-patient-out cross-validation (LOOCV). RESULTS: The subspace KNN classifier with 80-20 split validation yielded 90.4% accuracy with 92% sensitivity and 90.4% specificity (area under the curve [AUC] 0.968). The more stringent random forest with LOOCV approach achieved 81.1% accuracy with 71.5% sensitivity and 89.2% specificity (AUC 0.736). The device successfully distinguished between clean and infected CSF samples, with particularly high specificity in identifying noninfected samples. CONCLUSIONS: BOSoMetre shows promise as a low-cost (< €100), open-source tool for continuous CSF monitoring and early infection detection, especially for resource-limited settings. The high specificity could potentially reduce unnecessary CSF sampling and associated iatrogenic infection risks. Although the initial results are encouraging, further validation in larger cohorts is needed to confirm clinical utility and overcome the technical limitations identified in this proof-of-concept study.

来源:J Neurosurg,查看原文。版权归原作者所有。

引用

  1. Akbulut BB, Gürses BO, Özgül S, et al. Proof-of-concept study of noninvasive, rapid, machine learning-enhanced, color-based CSF diagnostics: a novel approach to external ventricular drain infection screening. J Neurosurg. 2025. PMID: 40972036. doi:10.3171/2025.5.JNS25628

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。