Alzheimer's disease (AD) and other dementias have become the fifth leading cause of death worldwide. Accurate early detection of the disease and its precursor, Mild Cognitive Impairment (MCI), is crucial to alleviate the burden on the healthcare system. While most of the existing work in the literature applied neural networks directly together with several data pre-processing techniques, we proposed in this paper a screening system that is to perform classification based on automatic processing of the transcripts of speeches from the subjects undertaking a neuropsychological test. Our system is also shown applicable to different datasets and languages, suggesting that our system holds a high potential to be deployed widely in hospitals across regions. We conducted comprehensive experiments on two different languages datasets, the Pitt dataset and the NTUHV dataset, to validate our study. The results showed that our proposed system significantly outperformed the previous works on both datasets, with the score of the area under the receiver operating characteristic curve (AUROC) of classifying AD and healthy control (HC) being as high as 0.92 on the Pitt dataset and 0.97 on the NTUHV dataset. The performance on classifying MCI and HC remained promising, with the AUROC being 0.83 on the Pitt dataset and 0.88 on the NTUHV dataset.

阿尔茨海默病 (AD) 和其他痴呆症已成为全球第五大死因。准确及早发现疾病及其前兆，轻度认知障碍 (MCI)，对于减轻医疗保健系统的负担至关重要。虽然文献中的大多数现有工作直接将神经网络与几种数据预处理技术结合使用，但我们在本文中提出了一种筛选系统，该系统将基于自动处理来自从事神经心理学的受试者的语音记录进行分类。测试。我们的系统还显示适用于不同的数据集和语言，这表明我们的系统具有在跨地区医院广泛部署的巨大潜力。我们对两种不同语言的数据集进行了综合实验，Pitt 数据集和 NTUHV 数据集，以验证我们的研究。结果表明，我们提出的系统在两个数据集上均显着优于之前的工作，在 Pitt 数据集上对 AD 和健康对照 (HC) 进行分类的接收者操作特征曲线 (AUROC) 下面积的得分高达 0.92，并且NTUHV 数据集上的 0.97。分类 MCI 和 HC 的性能仍然很有希望，AUROC 在 Pitt 数据集上为 0.83，在 NTUHV 数据集上为 0.88。97 在 NTUHV 数据集上。分类 MCI 和 HC 的性能仍然很有希望，AUROC 在 Pitt 数据集上为 0.83，在 NTUHV 数据集上为 0.88。97 在 NTUHV 数据集上。分类 MCI 和 HC 的性能仍然很有希望，AUROC 在 Pitt 数据集上为 0.83，在 NTUHV 数据集上为 0.88。