基於自傳式記憶測驗之雙模態縱向認知障礙檢測系統

Abstract

近年來，隨著人口老化現象顯著增長，認知疾病的發病率也隨之增加。在這一系列疾病中，阿茲海默症佔據了相當大的比例，對醫療系統造成了高成本的負擔。為了及早進行治療，延緩患者的退化過程，及時診斷出中間狀態的輕度認知障礙（MCI）是非常重要的。在本論文中，我們利用自傳式記憶測驗的語音資料，建立了一套針對MCI的雙模態縱向認知檢測系統。自傳式記憶測驗是一種評估受試者認知的心理學測驗模式，受試者在測驗過程中會自由講述其人生中的重要經歷。相比傳統測驗，在非結構化的自發性語音中找到隱含的疾病資訊更具挑戰性。在我們的研究中，我們從語音和文字的角度切入資料，提供更豐富的線索來判斷受試者的認知狀態。此外，為了捕捉自發性語音在不同時間點的認知變化，我們提出了老化軌跡模組計算局部與全局的對齊損失函數，通過對齊認知變化的方式增強時間變化上的特徵學習。在我們的中文數據集實驗中，包含老化軌跡模組的模型在多時間點數據集中的兩種資料上AUROC分別達到了85%和89%，相比單時間點的模型有了顯著的進步，同時我們也進行消融實驗，驗證提出了老化軌跡模組的必要性。為了證實模型不僅適用於自傳式記憶測驗資料，我們也將模型的一部分應用於單時間點的半結構化資料上進行驗證，結果顯示模型的正確率均超過78%。

In recent years, the incidence of cognitive diseases has also risen with the significant increase in population aging. Among these diseases, Alzheimer's disease constitutes a substantial proportion, placing a high-cost burden on healthcare systems. To give early treatment and slow the progression of patient deterioration, it is crucial to diagnose Mild Cognitive Impairment (MCI), a transitional stage. In this thesis, we utilize autobiographical memory (AM) test speech data to establish a dual-modal longitudinal cognitive detection system for MCI. The AM test is a psychological assessment method that evaluates the cognitive status of subjects as they freely narrate important life experiences. Identifying hidden disease information in unstructured, spontaneous speech is more challenging than traditional tests.

In our study, we analyze data from both speech and text perspectives, providing richer clues to assess the cognitive state of the subjects. Additionally, to capture cognitive changes in spontaneous speech over different time points, we propose an aging trajectory module that calculates local and global alignment loss functions, enhancing temporal feature learning by aligning cognitive changes over time. In our experiments on the Chinese dataset, the model incorporating the aging trajectory module achieved AUROC of 85% and 89% on two data, respectively, showing significant improvement over single time-point models. We also conducted ablation studies to verify the necessity of the proposed aging trajectory module. To confirm that the model not only applies to autobiographical memory test data, we used part of the model to single time-point semi-structured data for validation, with results showing an accuracy exceeding 78%.