Abstract: In order to objectively reflect the information carried by spectral data of tongue diagnosis of Traditional Chinese Medicine（TCM），near infrared spectroscopy and neural networks are firstly adopted to diagnose the disease.Three models，BP neural network，GRNN and PCA-GRNN，are employed for analysis and prediction of tongue spectral.Firstly the three modeling methods are analyzed，then spectral data from the healthy and diabetic patients are collected to correct tongue spectral model.Two kinds of samples including 39 cases of tongue spectrum respectively，a total of 78 samples，are to be divided into training set of 60 samples and prediction set of 18 in the neural network learning，and they are respectively trained and predicted under the three models.The result is that the smallest mean absolute error of PCA-GRNN of the three models is 13.2%，a minimum of training time 0.072 255 s，and in the condition that relative deviation is less than 0.5，the correct rate is 100%.It shows that the PCA-GRNN model can be applied to the analysis of tongue spectrometry and obtains better results，which is helpful for objectiveness of tongue diagnosis of TCM.

Key words: BP neural networks, general regression neural network, principal component analysis, tongue diagnosis, near infrared spectroscopy

摘要： 为了能客观地反映中医舌诊所蕴涵的病理信息，首次采用近红外光谱和神经网络对疾病进行诊断。分别采用了BP网络、广义回归神经网络（GRNN）、主成分分析和广义回归神经网络（PCA-GRNN）结合的三种模型在舌诊光谱法中的分析预测，首先对三种建模方法进行了分析，再用采集的健康人和糖尿病患者舌诊光谱数据进行校正模型的建立，两类舌诊光谱样本各39例，共计78例样本，在神经网络学习中，将其分成训练集样本60例和预测集样本18例，分别利用所建的三种模型对舌诊光谱样本进行训练和预测。实验结果是三种模型中PCA-GRNN相结合的方法平均绝对误差最小为13.2%、训练时间最短为0.072 255 s，以相对偏差在0.5以内为正确的情况下，其正确率为100%。说明用PCA-GRNN模型可以应用于舌诊光谱法的分析，并取得较好的分析结果，这对中医舌诊的客观化起到了一定的推动作用。

关键词: BP神经网络, 广义回归神经网络, 主成分分析, 舌诊, 近红外光谱