嵌入停頓編碼之具對比增強的認知障礙自動檢測系統

Abstract

隨著全球老年人口的逐年增長，阿爾茨海默病患者也同樣地增加，現有的醫療保健系統由於患者對治療和早期診斷的高需求導致相當大的負擔，因此，對認知障礙篩查系統的研究被廣泛關注以協助醫生診斷阿爾茨海默病去降低負擔。在本論文中，我們提出了一種基於自動轉錄的嵌入停頓編碼之具對比增強的認知障礙自動檢測系統。對於認知障礙，語音中的停頓模式是一種常用的聲學特徵，可以提供更多信息給予模型去進行更好的判斷，此外，反向翻譯和對比學習將使我們的對比增強模型在隱藏空間上有更好的表示，對比模型在微調具暫停嵌入的轉錄後可用來檢測患者的認知障礙。為了提高所提出之系統在現實世界中的適用性，我們的系統是全自動的，並且可以生成可以解釋的的結果，我們也使用英語和中文兩種語言評估我們的系統，兩種成功的結果都證明了我們系統的多語言能力。在對我們工作的定量評估方面，我們的系統在ADReSS數據集上自動檢測阿爾茨海默病可以達到81%的準確率，同時。此外，我們的系統在解決檢測輕度認知障礙（介於健康和阿爾茨海默氏症之間的中間階段）這一更具挑戰性的任務方面的準確性亦有不錯的表現。我們亦擴展檢測輕度認知障礙的任務到更非結構化語音的數據集，也就是我們於本地端收集的自傳式記憶數據集上，我們的系統的準確率平均可以達到71%準確率。

As the global elderly population grows annually, healthcare systems face a burden from the rise in Alzheimer's patients due to its high demand for treatment and early diagnosis. Therefore, research on cognitive impairment screening systems is studied widely to assist doctors in diagnosing Alzheimer's disease. In this thesis, we propose a contrast-enhanced automatic cognitive impairment screening system embedded with paused encoding based on automatic transcription. For cognitive impairment, the pause pattern in speech is a commonly studied acoustic feature that can provide more information based on which the model can make a better distinguishing judgment. Moreover, back-translation and contrastive learning represent a better contrast-enhanced model. After fine-tuning the transcripts embedded with pause, such a contrast-enhanced model is applied to detect the patients' cognitive impairment. To improve the applicability to the real world, our system is fully automatic, and its generated results can be shown to be explainable. We evaluate our system in two languages, English and Chinese, and both successful results demonstrate the multi-lingual ability of our work. In terms of quantitative evaluation of our work, our system can achieve 81% accuracy while automatically detecting Alzheimer's disease on the public ADReSS dataset. Besides, the accuracy of our system in tackling a more challenging task of detecting mild cognitive impairment (MCI), the middle stage between healthy and Alzheimer's, is highly promising. As for the same task of detecting MCI, on a more unstructured speech dataset, called autobiographical memory dataset collected locally, we show that the accuracy of our system can reach 71% on average.