基於獨白文字紀錄之失智症評估分類器

Abstract

這項研究是關於AD（阿茲海默症）的檢測，受到先前日本相關研究的啟發，在他們的研究中使用到了聊天機器人來陪伴老年人，並藉由穿插MMSE問題到日常對話中以利於評分老人目前的認知能力情況，此系統改善了老年人護理中短期追蹤的不利情況，雖然我們的研究並沒有用到聊天機器人，但是建構了一個完整的分類器與文字分析系統。 透過使用受測者的圖片獨白文字資料來完成AD的分類器，這裡有三個主要的貢獻如下，第一個是我們自己創建了的中文語料集(正常病患由我們自己收集，AD病患則是線上開源取得)。第二，我們提出一個新的特徵提取流程，它包括兩個主要部分，先驗特徵和抽象特徵，先驗特徵包括語法和語義特徵，這兩個都是藉由語言學上的知識求出。抽象特徵則是通過使用深度神經網絡技術，利用CNN和SPA建立一個端到端的模型來訓練求得。第三，對AD的特徵進行視覺化分析，並藉由多變量分析獲得每個先驗特徵之間的相關性。最後在比較結果的同時，也有與2017年的日本研究做相比，確實有較好的表現。

This research is about the AD (Alzheimer’s disease) detection. It is inspired by a former research in Japan, that they use a chatbot to keep elderlies company. The daily conversation is added with the MMSE questions to get the score of the elderly. This system improves the situation of short-term chasing in the elderly cares. Although we do not create a chatbot, a complete system of text analysis for the AD picture monologue data is built. A classifier of AD is implemented by using the subject’s picture monologue language data. Here come three main contributions as follows. The first one is the Chinese language data set are created by ourselves. The second one is we propose a new process of feature extraction, which consists of two main part, i.e., prior features and abstract features. Prior features consist of syntactic and semantic features by linguistic knowledge. Abstract features are acquired by using the deep neural network techniques, CNN and semantic pointer architecture, and training on an end-to-end model. The third one is making a visualization analysis on the features of AD and control normal to get the underlying knowledge. The multivariate analysis is also done and gets the correlation coefficients of each feature-pairs. The results have been compared with the former works and get a better performance.