利用機器學習方法分析電子氣體感測資料以鑑別慢性肺阻塞與氣喘患者

Discrimination of Patients with Chronic Obstructive Pulmonary Disease and Bronchial Asthma by Applying Machine Learning Methods to analyze Electronic Gas-Sensing Data

李宥瑾(Li,You Jin)

指導教授 ： 劉奕汶

國立清華大學/電機資訊學院/電機工程學系/碩士(2015年)

https://doi.org/10.6843/NTHU.2015.00421

摘要

利用人工嗅覺的技術協助醫生了解病患、輔助醫生判斷病患的生理狀況，是目前很有潛力的生醫電子應用。由於氣喘與慢性肺阻塞兩種疾病之病徵極為相似，若能夠早期偵測、區分，在臨床上將有應用的價值。本研究針對這個主題，採用了化學感測器陣列與患者呼出之氣體反應，取得了反應訊號後，首先藉由訊號前處理技術，消除基線飄移及樣本間的濃度差異。接下來，氣體的辨識流程分為以下幾個部分：一、先行辨識該氣體是否為病患所呼出；二、若辨識結果為病人氣體，則分析該名患者罹患何種疾病(氣喘或慢性肺阻塞)；最後，為該名病患所罹患的疾病分析其患病嚴重度。辨識方法則是利用主成分分析法結合線性判別分析降低資料維度，減少資料點運算量；再使用支持向量機或K個最近鄰居法兩者相互比較，再以混淆矩陣觀察之。除了以上辨識分析外，我們更進一步地觀察主成分分析法降維時的各感測器權重，並以感測器之目標氣體與氣喘及慢性肺阻塞兩種患者呼出的常見揮發性有機化合物比較之。本篇論文的辨識結果發現，經由以上流程之辨識結果均達八成以上，且當我們在辨識疾病以主成分分析法回推降維的感測器權重時，發現其中一個權重較高的感測器TGS2620，該感測器之目標氣體二甲苯，是分辨慢性肺阻塞患者與氣喘患者呼出的氣體中，明顯不同的揮發性化合物之一。

關鍵字

機器學習方法 ； 氣體感測 ； 慢性肺阻塞 ； 氣喘

並列摘要

The use of artificial olfactory technology to assist medical diagnosis is a promising domain for electronic applications. Since bronchial asthma (BA) and chronic obstructive pulmonary disease (COPD) have very similar symptoms, being able to detect and discriminate them automatically would have positive impact in the clinics. In this research, we focused on this topic and used a chemical sensing array to detect the gas exhaled from the patients. Signal preprocessing techniques were applied to remove the baseline drift and normalize the influence of concentration between samples. Then, methods for gas classification involved the following steps: first, we determined whether the gas was produced by a patient who suffered from either kind of the diseases. Secondly, if the classification result was positive, we applied further analysis to tell what kind of disease the patient had. Finally, we analyzed the severity of the patients. The method of classification was based on principal component analysis (PCA) and linear discriminant analysis to reduce the dimension of data. After that, we compared the result of the support vector machine and the K nearest neighbor method to achieve the best performance. We used the receiver operating curve and the area under the curve to assess the reliability and validity of the classifier. In addition, we observed the weights of the sensors in reduced dimensions of PCA and checked whether the main components of the analyzed result agreed with the list of commonly observed volatile organic compounds in the gas exhaled by patients of COPD and BA. The results of the identification all have a recognition rate of above 80%. Also, PCA-based analyses indicate that xylene, which is one of the volatile organic compounds found in gas exhaled by COPD patients, is a key compound that enable discrimination of COPD and BA.

並列關鍵字

Machine Learning Methods ； Electronic Gas-Sensing ； Chronic Obstructive Pulmonary Disease ； Bronchial Asthma

參考文獻

[11] 楊廷然, 「利用多標籤分類器實現電子鼻混合氣體識別方法之研究」, 國立清華大學電機工程研究所, 碩士論文, 2014.

[13] 許柏安, 「快速混合氣體辨識方法之研究」, 國立清華大學資訊工程研究所, 碩士論文, 2012.

[14]王家銘, 「利用樣式識別實現電子鼻肺炎偵測」, 國立清華大學電機工程研究所, 碩士論文, 2013.

[2] W. Wilkens and J. Hartman, “An electronic analog for the olfactory processesa,” Journal of Food Science, 1964.

[3] T. Pearce, S. Schiffman, H. Nagle, and J. Gardner, “Handbook of machine olfaction: electronic nose technology,” 2006.