Helgason C M, Malik D S, Cheng S C, Jobe T H, Mordeson J N
Department of Neurology and Psychiatry, University of Illinois College of Medicine at Chicago, Ill., 60611, USA.
Neuroepidemiology. 2001 May;20(2):77-84. doi: 10.1159/000054764.
Evidence-based medicine, founded in probability-based statistics, applies what is the case for the collective to the individual patient. An intuitive approach, however, would define structure in the (physiologic) system of interest, the human being, directly relevant to other systems (patients) composed of similar variables. A difference in measure of variable interaction in the patient from that in the collective would show how extrapolation of information from the latter to the single patient is counterintuitive.
We compare statistical to 'fuzzy' measures of variable interaction. Three diagnostic variables are considered in 30 stroke patients who underwent the same diagnostic tests. 'Fit' (fuzzy information) values [0, 1] for degree of variable severity were expertly assigned by 2 blinded raters for real and fabricated patients. Fabricated patients were composed of real-patient 'fit' values after shuffling. Real and fabricated patients were each numerically represented as a set. Three groups of fabricated patients and the real patient group were studied. Statistical [Pearson's product-moment (regression analysis) and Spearman's rank correlation] and three different fuzzy measures of variable interaction were applied to patient data.
Interaction for blood-vessel measured strong in real patients, and weak after one shuffle, using all fuzzy measures. By comparison, the same interaction was found in real patients by only 1 rater (Rater 2) using 1 statistical technique (Spearman's rank correlation) which, as did Pearson product-moment correlation, found a 'significant' interaction between blood-heart in fabricated patients.
Our study suggests that the measure of variable interaction in nature - as combined in the individual (real) patient - is captured robustly by fuzzy measures and not so by standard statistical measures.
循证医学建立在基于概率的统计学基础上，将适用于总体的情况应用于个体患者。然而，一种直观的方法会直接在感兴趣的（生理）系统（即人类）中定义结构，该系统与由相似变量组成的其他系统（患者）直接相关。患者中变量相互作用的测量结果与总体中的测量结果存在差异，这将表明从总体到单个患者的信息外推是违反直觉的。
我们比较了变量相互作用的统计测量方法和“模糊”测量方法。在30名接受相同诊断测试的中风患者中考虑了三个诊断变量。两名不知情的评估者为真实患者和虚拟患者专业地分配了变量严重程度的“拟合”（模糊信息）值[0, 1]。虚拟患者由真实患者的“拟合”值洗牌后组成。真实患者和虚拟患者分别用一组数字表示。研究了三组虚拟患者和真实患者组。将统计方法[皮尔逊积矩（回归分析）和斯皮尔曼等级相关]以及三种不同的变量相互作用模糊测量方法应用于患者数据。
使用所有模糊测量方法，在真实患者中血管相互作用测量结果很强，而在一次洗牌后变弱。相比之下，只有一名评估者（评估者2）使用一种统计技术（斯皮尔曼等级相关）在真实患者中发现了相同的相互作用，该技术与皮尔逊积矩相关一样，在虚拟患者中发现了心脏与血液之间的“显著”相互作用。
我们的研究表明，自然界中变量相互作用的测量——如在个体（真实）患者中组合的那样——通过模糊测量方法能够有力地捕捉到，而标准统计测量方法则不然。
