| 摘要: | 失語症是一種嚴重影響患者語言能力的疾病。 主要原因通常是中風，但也可能由其他因素引起，例如腦外傷、腫瘤和退化。 患者最常見的臨床表現包括找詞困難、導致言語錯誤、用其他詞替換目標詞以及聽覺理解、閱讀和寫作方面的困難。失語症的常見診斷方法包括臨床量表和非侵入性腦刺激，例如重複經顱磁刺激（rTMS）結合強化語言訓練。 前一種方法比較主觀，需要語言治療師的專業評估，而後者則費用更高。 因此，有效且客觀的診斷方法對於語言治療至關重要。本研究提出了一種評估失語症的自動化方法。 本實驗採用機器學習（ML）設計自動評估的算法模型，結合自主研發的VR語言訓練模塊，從任務執行中獲取行為和生理信息。 通過機器學習分析患者在各種語言任務訓練中的表現和情況。 研究結果將從統計分析和機器學習兩個方面進行討論。在統計分析中，我們將對正常個體和失語症患者之間的多模態生理和遊戲任務特徵進行Mann-Whitney U檢驗，與正常個體相比顯示出許多顯著差異（p < 0.05）。 在機器學習方面，評估結果表現良好，所有模型都達到了80%以上的準確率。;Aphasia is a disease that can seriously affect the language abilities of patients. The primary cause is usually stroke, but it can also be caused by other factors such as brain trauma, tumors, and degeneration. The most common clinical manifestations in patients include difficulty finding words, leading to speech errors, substituting target words with other words, and difficulties with auditory comprehension, reading, and writing.Common diagnostic methods for aphasia include clinical scales and non-invasive brain stimulation, such as repetitive transcranial magnetic stimulation (rTMS), combined with intensive language training. The former method is more subjective and requires professional evaluation from a speech therapist, while the latter is more expensive. Therefore, an effective and objective diagnostic method is crucial for language therapy.This study proposes an automated method for evaluating aphasia. In this experiment, machine learning (ML) is used to design an algorithm model for automatic assessment, combining self-developed VR language training modules to obtain behavioral and physiological information from task execution. The performance and situations of patients in various language task training are analyzed through machine learning. The research results will be discussed from two aspects: statistical analysis and machine learning.In the statistical analysis, we will conduct a Mann-Whitney U test on the multimodal physiological and game task features between normal individuals and aphasia patients, showing many significant differences compared to normal individuals (p < 0.05). In the machine learning aspect, the evaluation results show good performance, with all models achieving over 80% accuracy. |
