LEARNING FOR PATTERN CLASSIFICATION USING MULTI-OBJECTIVE PROGRAMMING AND ITS APPLICATON TO DIAGNOSIS SUPPORT SYSTEM OF DIABETIC ANGIOATHY

为了构建糖尿病血管病变诊断支持系统，我们对899例日本NIDDM合并和不合并大血管病变的大血管病变危险因素的特征进行了研究。来自40个机构的糖尿病大血管病变组的多临床研究登记。386名受试者被确认患有大血管病变(MA(+))；其中包括217名缺血性心脏病，169名脑血管疾病和77名外周血管疾病。单因素和多因素分析显示年龄、空腹血糖、高血压、收缩压、舒张压、糖尿病病程、糖尿病微血管病变、高密度脂蛋白胆固醇、低密度脂蛋白胆固醇、低密度脂蛋白胆固醇/低密度脂蛋白胆固醇、低密度脂蛋白胆固醇/低密度脂蛋白胆固醇、低密度脂蛋白胆固醇/低密度脂蛋白胆固醇、低密度脂蛋白胆固醇/低密度脂蛋白胆固醇比值、吸烟指数、体重指数是影响MA(+)、IHD、CVD、PVD的主要因素。综上所述，在NIDDM患者中，年龄、高血压、收缩压、舒张压和糖尿病病程是大血管病变的危险因素。我们开发了两种模式分类方法。其中之一是使用模糊和/或多目标线性规划来获得分段线性判别函数的方法。另一种是委员会机器，它是一个由多个子模块神经网络组成的复杂神经网络。已经观察到，这两种方法都可以有效地应用于我们的诊断问题。

In order to construct a diagnosis support system of diabetic angiopathy, we examined characteristic features of risk factors for macroangiopathy in 899 Japanese NIDDM with and without macroangiopathy. They were registered from 40 facilities by Multiclinical Study for Diabetic Macroangiopathy group. Three hundred eighty six subjects were identified as having macroangiopathy (MA (+) total) ; these includes 217 with ischemic heart disease (IHD), 169 with cerebrovascular disease (CVD), and 77 with peripheral vascular disease (PVD). Univariate and multivariate analyzes revealed the following factors for MA (+) total, IHD,CVD and PVD : age, fast blood sugar, hypertension, systolic blood pressure, diastolic blood pressure, duration of diabetes, diabetic microangiopathy, high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), LDL-C : HDL-C ratio, brinkman index (smoking) and body mass index. In conclusion, in NIDDM patients, age, hypertension, systolic blood pressure, diastolic blood pressure and duration of diabetes were found to be risk factors for macroangiopathy.Technologies of machine learning is applied for supporting diagnosis. We developed two kinds of methods for pattern classification. One of them is a method for getting a piecewise linear discrimination function using fuzzy and/or multi-objective linear programming. The other is a committee machine which is a complex neural network consisting of several submodular neural networks. It has been observed that both methods can be effectively applied to our diagnosis problem.