Skip to main content
QUICK REVIEW

[论文解读] Alternating Decision trees for early diagnosis of dengue fever

M. Naresh Kumar|arXiv (Cornell University)|May 31, 2013
Mosquito-borne diseases and control参考文献 16被引用 10
一句话总结

本研究提出一种结合提升法的交替决策树(ADTree),利用基础临床与实验室特征实现登革热的早期、精准诊断。该方法分类准确率达到89%,F1值为0.873,曲线下面积(AUC)为0.826,优于C4.5(准确率78%,AUC 0.617),表明其在资源有限的登革热监测中具有潜力。

ABSTRACT

Dengue fever is a flu-like illness spread by the bite of an infected mosquito which is fast emerging as a major health problem. Timely and cost effective diagnosis using clinical and laboratory features would reduce the mortality rates besides providing better grounds for clinical management and disease surveillance. We wish to develop a robust and effective decision tree based approach for predicting dengue disease. Our analysis is based on the clinical characteristics and laboratory measurements of the diseased individuals. We have developed and trained an alternating decision tree with boosting and compared its performance with C4.5 algorithm for dengue disease diagnosis. Of the 65 patient records a diagnosis establishes that 53 individuals have been confirmed to have dengue fever. An alternating decision tree based algorithm was able to differentiate the dengue fever using the clinical and laboratory data with number of correctly classified instances as 89%, F-measure of 0.86 and receiver operator characteristics (ROC) of 0.826 as compared to C4.5 having correctly classified instances as 78%,h F-measure of 0.738 and ROC of 0.617 respectively. Alternating decision tree based approach with boosting has been able to predict dengue fever with a higher degree of accuracy than C4.5 based decision tree using simple clinical and laboratory features. Further analysis on larger data sets is required to improve the sensitivity and specificity of the alternating decision trees.

研究动机与目标

  • 开发一种稳健且可解释的决策树模型,利用常规可获得的临床与实验室特征实现登革热的早期诊断。
  • 解决当前世卫组织分类方法的局限性,该方法依赖于早期登革热阶段并不存在的症状。
  • 相比传统决策树(如C4.5),提高诊断准确率并减少假阳性结果。
  • 在缺乏先进实验室检测的资源有限环境中,实现成本效益高的早期分诊。
  • 验证提升法在提升决策树模型预测性能方面对登革热分类的有效性。

提出的方法

  • 应用结合AdaBoost的交替决策树(ADTree),将弱预测规则整合为强集成分类器。
  • 采用10折交叉验证评估模型性能,防止过拟合。
  • 对连续变量进行预处理,缺失值以均值替代;因缺失率超过40%,排除IgM和IgG指标。
  • 将脉搏变量离散化为“低(L)”和“高(H)”两类,以提升分类准确率。
  • 通过多项逻辑回归(Wald检验)和卡方检验筛选显著特征。
  • 使用Weka 3.6.1训练并评估模型,性能通过混淆矩阵、F1值和ROC曲线进行评估。

实验结果

研究问题

  • RQ1与传统决策树(如C4.5)相比,结合提升法的交替决策树是否能利用基础临床与实验室数据提升早期登革热诊断的准确率?
  • RQ2通过统计与机器学习方法分析,哪些临床与实验室特征对登革热最具预测性?
  • RQ3对脉搏变量进行离散化如何影响ADTree模型的分类性能?
  • RQ4ADTree模型在早期登革热检测中相比C4.5,能多大程度上减少假阳性诊断?
  • RQ5在有限数据(n=65)上训练的模型是否能达到与在更大数据集上训练的模型相当的性能?

主要发现

  • 在对脉搏变量进行离散化后,ADTree模型的分类准确率达到89%,显著优于C4.5的78%准确率。
  • ADTree的F1值(0.873)和受试者工作特征曲线下面积(AUC 0.826)均高于C4.5的F1值(0.738)和AUC(0.617)。
  • ADTree正确分类了全部53例登革热阳性病例(真阳性率为100%),而C4.5漏诊了3例,表明其敏感性更优。
  • ADTree的假阳性率(58.3%)低于C4.5(91.6%),表明其在非登革热病例中具有更好的特异性。
  • 尽管数据量有限(n=65),模型性能依然稳健,敏感性达89.2%,特异性为52.4%。
  • 对脉搏变量的离散化是提升ADTree准确率的关键因素,凸显了在临床机器学习模型中特征工程的重要性。

更好的研究,从现在开始

从阅读论文到最终审阅,大幅缩短您的研究时间。

无需绑定信用卡

本解读由 AI 生成,并经人工编辑审核。