運用機器學習模型最佳化偵測輕度認知缺損患者中的認知與主觀認知抱怨型態

Optimizing Machine Learning Models for Detecting Cognitive Changes and Subjective Cognitive Complaint Patterns among Patients with Mild Cognitive Impairment

早期偵測阿茲海默病（Alzheimer’s disease, AD）患者的認知症狀以及判定患者的認知狀態為重要工作。近期研究顯示機器學習演算法對於促進AD患者的早期檢測具有潛在價值。然而，很少有研究使用機器學習演算法探討神經心理功能評估在偵測和判定AD患者認知狀態中的價值。本研究納入198名個體，包括認知未受損組（cognitively unimpaired, CU, n = 30）、輕度認知受損組（mild cognitive impairment, MCI, n = 125），以及一組患有輕度阿茲海默型失智症的患者（dementia of Alzheimer’s type, DAT, n = 43）、比較ML演算法，與傳統羅吉斯迴歸方法利用神經心理功能以及人口統計學和臨床特徵辨別不同認知狀態個體的能力，結果顯示，整體而言，ML法在準確率和操作特徵曲線下之面積的表現優於優於傳統羅吉斯模型。進一步地ML法在判定CU和MCI方面優於傳統羅吉斯迴歸模型，而在判定DAT方面ML法與傳統羅吉斯迴歸模型的表現相近；在對於認知狀態的辨別上，神經心理功能相關表現相對其他面向訊息大多有較高的貢獻度。據我們所知，這是台灣第一個將機器學習與神經心理功能評估結合的研究。機器學習演算法可能有助於早期發現和區分AD患者的認知症狀。

Early detecting cognitive symptoms and determination of cognitive status among patients with Alzheimer’s disease (AD) are important works. Recent research has demonstrated the potential values in facilitating the early detection of AD. However, very few studies investigated the values of performance on neuropsychological assessment in the detection and determination of cognitive status among AD patients with machine learning (ML) algorithms. The present study compared the discriminative abilities of neuropsychological functions and demographic and clinical features among patients using ML and traditional logistic regression approaches among 198 individuals comprising a cognitively unimpaired group (CU, n = 30), a mild cognitive impairment group (MCI, n = 125), and a group of patients with mild dementia of Alzheimer’s type (DAT, n = 43). Results revealed, in general, the ML method outperformed the traditional logistic regression model in terms of accuracy and area under the operating characteristic curve. Furthermore, the ML method is better than the traditional logistic regression model in classifying CU and MCI, while the performance of the ML method is similar to the traditional logistic regression model in classifying DAT. To the best of our knowledge, this is the first study incorporated ML and neuropsychological function assessment in Taiwan. ML algorithms might have values in facilitating the early detection and differentiation of cognitive symptoms among patients with AD.