A Digital Auscultation System For Five Organ Illness Diagnosis In Traditional Chinese Medicine

Auscultation is an important diagnosis method in traditional Chinese medicine (TCM) to determine the patients’physiological condition and pathology by detecting their vocal changes. Traditional auscultation of TCM mainly depends on the auditory sense of experienced Chinese physician. The diagnosis is not objective enough due to the difference of clinical settings, physicians’experience and sensibility of hearing. The assessment lacks a unified and quantified standard. A digitized auscultation system is an intelligent diagnosis apparatus built on computation devices (e.g a computer) which can give quantified, objective and automatic diagnosis for patients by collecting and analyzing their speech voice.This thesis is focused on the diagnosis of five organ diseases. Five organ diagnoses is an auscultation method recorded in the earliest TCM classics with a long history dating back to2000years ago. However it is rarely conducted in clinical practice nowadays. According to the theory of TCM, there are5tones corresponding to5organs in human body. By listening to the changes in the5tones pronounced by the patient, the physician can determine the functional activity of the5organs. Therefore in this research, we collected the phonation of ten Chinese words associated with the five tones for analysis. Two kinds of feature extraction methods i.e. the Mel-frequency cepstrum coefficient (MFCC) and the Bark SampEn are proposed and found to be effective in discriminating the6types of subjects we have surveyed. In further investigation, we conducted a feature selection procedure on the parameters we have extracted. For both the MFCC and the Bark parameters, an optimal feature set is formed that can dramatically reduce the dimension of the feature space and improve the efficiency of classification task.Besides the feature extraction and feature selection procedure, a successful auscultation system needs a good classifier to recognize the type of disease on the basis of the extracted feature. Support vector machine (SVM) is utilized in this thesis for classification task. Considering the problems involved in this research are mostly multi-category classification tasks, we proposed two approaches to improve the performance of the traditional multi-class SVM (MSVM) algorithm. The experiments on public datasets and the auscultation datasets clearly verify the effectiveness of our improved version of MSVM. With these improvements, we achieved75%correct rate on five-organ diagnosis and80%accuracy on qi/yin deficiency diagnosis.In view of the multi-syndrome cases in clinical practice, we studied the qi/yin deficiency diagnosis using multi-instance multi-label (MIML) learning approach. The experiment data are the patients’speech samples of5vowels i.e./a/,/o/,/e/,/i/,/u/. Each patient in the dataset may have one or both of the qi and yin deficiency syndromes. By regarding the5vowel samples from one patient as instances and the patient’s syndrome type as the labels, the problem can be properly formalized under multi-instance multi-learning framework.