基於情緒感知的長者認知功能障礙偵測

Abstract

隨著全球人口的高齡化，失智症，這種因異常腦部變化導致並好發於長者的症候群，已成為社會中的重大議題。輕度認知障礙作為介於正常老化與失智症之間的階段，及早檢測可以提供即時的介入治療，從而提高患者的生活品質並可能防止其進一步惡化。為此，我們提出了一種基於情緒感知的認知功能障礙檢測方法。為了實現這一方法，我們建立了一個人機互動框架，包括人臉辨識系統、多語言對話系統、情緒辨識系統以及我們實驗室之前開發的運動輔助系統。我們提出的認知功能障礙檢測方法利用從對話和運動人機互動中收集的數據來分類三種認知能力階段:正常、輕度認知障礙和高度認知障礙。 為了評估此方法的可行性和有效性，我們在桃園龜山的大崗、嶺頂與陸光社區招募了七十名年齡介於58至97歲之間的受試者。在實驗中，受試者與機器人進行論文中提出的對話和運動互動，並利用蒙特利爾認知評估-台灣版 (MoCA-T)來評估受試者的認知功能。我們將根據MoCA-T結果分類出的認知能力階段作為基準真相 (ground truth)，並使用支援向量機 (Support Vector Machine)作為我們都分類模型。實驗結果顯示，我們的方法在分類認知能力階段方面達到了92.3%的準確率。

With the global population aging, dementia, a syndrome caused by abnormal brain changes and prevalent among older people, has become a significant issue in society. Mild cognitive impairment (MCI) as the stage between normal aging and dementia. Ear-ly detection of cognitive impairment allows for timely intervention and treatment, im-proving the quality of life for individuals and potentially preventing its progression. Therefore, we proposed a cognitive function impairment detection method based on emotion perception. To implement this method, we establish a human-robot interaction framework consisting of a face recognition system, a multilingual dialogue system, an emotion recognition system, and an exercise assistance system previously developed by our laboratory. Our proposed cognitive function impairment detection method utilizes the data collected from the dialogue and exercise human-robot interactions to classify three cognitive function levels: normal, mild cognitive impairment, and highly cognitive impairment. To evaluate the feasibility and efficacy of our method, we recruited seventy participants aged 58 to 97 in Dagang, Lingding, and Luguang communities of Guishan, Taoyuan. In the experiment, participants engaged in both dialogue and exercise interactions with the robot. The Taiwan version of Montreal Cognitive Assessment (MoCA-T) was utilized to evaluate the cognitive function of the participants. Ground truth cognitive function levels were categorized based on the MoCA-T results. We employed Support Vector Machine (SVM) as our classification model. As the results, our method achieved 92.3% accuracy in classifying cognitive levels.