以機器學習方法利用語音資料對帕金森氏症和阿茲海默症患者進行病情偵測

Abstract

帕金森氏症和阿茲海默症屬於神經退化性疾病，目前在臨床診斷中沒有良好的預防措施以及不能完全的治療，一般情況中，醫生會依據患者自身的病症開出相對應的藥單，以延緩疾病的惡化，因此早期發現病情變得非常重要。然而，患者經常不知道或不願意承認自己得病，以至於延誤就醫導致病情日漸嚴重，這篇研究的目的是設計一個病情早期偵測的機器學習模型。 本研究提出以語音特徵進行病情早期偵測的機器學習系統。訓練資料集的收案地點為臺灣大學神經內科林靜嫻醫生和邱銘章醫生的門診。在帕金森氏症的研究中，資料有215名帕金森氏症患者和289名正常人，錄音內容來自於短篇文章的朗讀。在阿茲海默症的研究中，資料有68名阿茲海默症患者和23名正常人，錄音內容除了短篇文章還加入餅乾小偷的敘述錄音。聲學和自然語言處理的特徵提取中，不同於過往文獻選用自動化提取特徵模型，而是考慮病人在初期會出現的語音症狀，例如語調變化減少和音量會有漸小的情況，來進行特徵提取並在AI模型上進行訓練。 透過特徵選取後的特徵集對模型做訓練，並利用解釋AI的方式分析特徵在模型預測中影響的情況。實驗結果中包括五種的檢測評分來佐證模型預測的成效，還有利用特徵選取與解釋AI得來的特徵重要性排序，進一步比較在帕金森氏症和阿茲海默症的模型預測中特徵影響的不同。帕金森氏症檢測達到80%的招回率(recall rate)，阿茲海默症檢測則達到98%的招回率。

Parkinson’s disease (PD) and Alzheimer’s disease (AD) are progressive neurodegenerative disorder that has no known cure and no known prevention. Doctors prescribe drugs according to the patient’s symptoms to delay the disease's progression in clinical diagnosis. However, patients often don’t know or think they are sick. For this reason, early detection is crucial to slow down the progress. PD data provide 215 patients and 289 controls with the recording of word fluency tasks. AD data provide 68 patients and 23 controls with the recording of word fluency and open description tasks. Feature extraction by applying acoustic and text-based features is essential for disease detection. In the data preprocessing, experimental paradigm, normalization, impute missing value, and feature correlation is also provided. The common classifier, such as support vector machine (SVM), Decision tree, and random forest (RF), is used as the machine learning algorithm using the sequential feature selector method. The models are trained by a subset of input features from the dataset and analyzed by the explainable AI (XAI) method. The experimental results show the five evaluation scores and comparison of the feature importance sequence obtained in PD and AD research. The binary classification performance achieves over 80% recall in PD detection and over 98% in AD detection.