Alzheimer's disease (AD) is a neurodegenerative and progressive disease that cannot be cured effectively (1). Mild cognitive impairment (MCI) is the early stage of AD. The study by the Lancet Public Health in 2020 found that the prevalence of dementia in people over 60 years old in China accounted for 6.04% of the population (approximately 1,507 ten thousand), and the number was 15.54% (3,877 ten thousand) for MCI cases (2). An epidemiological survey also found that a person's cost with AD in China was approximately $19,144.36 in 2015, while the total cost of the world's average level was $167,740 million, which was composed of $54,530 million (32.51%) direct medical cost, $26,200 million (15.62%) direct non-medical cost, and $87,010 billion (51.87%) indirect cost (3).

In the past 20 years, scholars have reported extensive studies on the relationship between the pathogenesis of AD and language fluency (4). They generally believed that mild word naming, retelling, hearing, understanding, and writing disorders already exist in the early stages of AD. One of the early signs of AD is an obvious decline in linguistic comprehension and expression form (5), and the linguistic manifestation of patients with AD usually includes the following:

New studies have found that before the onset of AD, the β-amyloid has already gathered in the brain about 5 to 10 and even 20 years ago. If AD can be diagnosed at an early stage (6), a series of behavioral therapies can be prescribed to slow the progress of the disease. However, an AD diagnosis is challenging in clinical medicine because of the subtle differences between patients with AD and healthy individuals in terms of brain structure and behavior. At present, some medical diagnosis methods, such as pathological examination, MRI, PET, and reliable biomarkers (e.g., amyloid ligand imaging and cerebrospinal fluid testing), are usually used. However, these diagnosis methods cannot be widely popularized because of their high cost and invasive nature. Therefore, there is an urgent need to develop a convenient, inexpensive, and non-invasive AD diagnostic approach by AI technologies, such as speech processing and NLP. In contrast to earlier studies with manual expert-wise feature extraction in this field, this study used a reliable deep learning model to automatically find suspicious AD symptom features from speeches. Specifically, a pre-trained distilBert language model (7) was used as a feature extractor to obtain the features of the input sentence or document, and a simple logistic regression classifier, which has a good effect on binary classification, was used to classify AD from normal controls. Owing to its strong deep semantic feature extraction competency and an accurate binary classifier, this combination can effectively improve the classification effect. In addition, a grid search strategy (8) was used to tune the parameters to obtain the best parameters of the model. The results show that this method worked better on ADReSS datasets (9) in 2020, with an accuracy of 0.88, which was significantly higher than the baseline and almost equivalent to the best performance on the challenge (10).

The main contributions of this study are as follows:

Different technologies can be used to detect AD, such as molecular biomarkers combined with deep learning on gene expression datasets (11). However, we used transcripts combined with deep learning on speech datasets instead. Two approaches are mainly used in this field: machine learning with manual feature extraction based on expert knowledge and deep learning. Traditional machine learning algorithms have been widely studied with handcrafted features to predict AD. However, they have the disadvantage of lacking integrity, demanding good expertise, low accuracy, and poor portability. Moreover, these methods are generally applicable to a specific task scene. Once the scene changes, these manually designed features, and prior settings cannot be adapted to new scenes and need to be redesigned again; therefore, the portability of the model is not better overall. With the arrival of the deep learning paradigm, it has already become possible to extract high-level abstract features directly from transcripts that describe the distribution of datasets in low-dimensional manifolds internally. The advantage is that it can either extract input dataset patterns directly for both regression tasks or combine handcrafted features to the feature map of the input dataset without the certified professionals from the data source. Because language functions play an important role in the detection of cognitive deficits at different stages, the combination of NLP technology and deep learning provides an accurate and convenient solution for the detection of AD and MCI (12). In this study, a distilBERT model, which is a multi-layer perceptron with a self-attention mechanism, is used to extract deep semantic features; they are then passed through a strong binary classifier to recognize AD. The number of hidden layers is larger than that of traditional machine learning algorithms, thus, the model has a stronger semantic abstraction ability and classification performance, and the scalability is superior to traditional machine learning methods. Although the deep learning method does not need to extract features manually, it does not mean that we do not need to analyze manual features anymore, and the single deep learning model for diagnosing AD may perform better. Therefore, combining it with some conspicuous markers and a stronger classifier may improve the classification results, which we will discuss in the discussion section.

Several studies have investigated language and speech features for AD diagnosis (13) and proposed many signal processing and machine learning algorithms to detect AD and MCI (14). However, in this field, there are still lacking benchmark datasets against which different methods can be systematically compared. The ADReSS Challenge (9), a subset of the DementiaBank dataset (15), uses a balanced dataset of AD and healthy controls to recognize the disease. Manual feature extraction methods have a better interpretation for classification tasks, although there are unremarkable results. As a basic study on the ADReSS dataset, Luz et al. (9) used 34 linguistic features, such as total utterances, a type-token ratio, percentages of parts of speech, duration, MLU, and a word ratio, combined with linear discriminant analysis, and obtained the best accuracy of 0.75 on the test dataset. Acoustic features, such as emobase (16), the extended Geneva minimalistic acoustic parameter set (eGeMAPS) (17), minimal features (18), Computational Paralinguistics ChallengE (ComParE), (19), and multi-resolution cochleagram (MRCG) (20), only obtained an accuracy of approximately 0.5 on the classifiers used frequently. Balagopalan et al. (21) used two approaches for the binary classification of AD and normal controls, i.e., acoustic and text-based feature extraction and the bidirectional encoder representation (BERT) model. Finally, the BERT model obtained the best accuracy of 0.8332, which was better than that of the manual feature extraction method. Syed et al. (22) and Yuan et al. (23) achieved accuracies of 85.45 and 89.6% using acoustic and linguistic features, respectively. Syed et al. (22) used acoustic features, such as bag-of-acoustic-words and INTERSPEECH 2010 Paralinguistic Challenge feature sets [a low-dimensional version of ComParE (19)], and obtained an accuracy of 76.85%. Luz et al. (24) used a combination of phonetic and linguistic features without human intervention and obtained an accuracy of 78.87%. Most of these earlier studies were based on features designed by experts and were unable to learn more informative and discriminative features, so a relatively poor performance was obtained.

The latest deep-learning methods, such as convolutional neural networks (CNN), recurrent neural networks (RNN), and BERT, can achieve good performances by automatically extracting high-level features. Mahajan et al. (25) used part-of-speech (POS) tags and word embeddings (GloVe) as inputs on a CNN-long short-term memory (LSTM) model (26) and obtained the best accuracy of 0.6875. Then, they replaced unidirectional LSTM with bidirectional LSTM layers (27) and obtained the best accuracy of 0.7292. Orimaye et al. (28) used a deep neural network to predict MCI in speech. Different from our datasets, they used part of the Pitt corpus of the DementiaBank dataset, comprising 19 controls and 19 MCI transcripts. Fritsch et al. (29) enhanced n-gram language models to create neural network models with LSTM cells, and an accuracy of 85.6% was obtained to classify HCs and AD on the Pitt dataset. Pan et al. (30) used a glove word embedding sequence as the input, combined with gated recurrent unit layers and a stacked bidirectional LSTM to diagnose AD on the Pitt dataset. These models differ from our model because we used deep learning and machine learning classifiers instead. Similar to our method, the study (31) demonstrated that the combination of BERT_Large and logistic regression had the best performance in the classification problem. Different from our study, they used the Pitt DementiaBank dataset and data augmentation technology to enhance the classification performance and obtained a state-of-the-art (SOTA) accuracy of 88.08%.

Other tasks, except for the picture description task, can also be used to recognize AD. For example, Clarke et al. (32) used five different tasks to recognize AD, namely, conversation, procedural recall, picture description, narrative recall, and novel narrative retelling and obtained the best accuracy of 90% for AD vs. HC with linguistic features, combined with the support vector machine (SVM) model. In addition, many studies have used multimodal datasets to detect AD and MCI, and more accurate and differentiated information may be obtained from different models. Looze et al. (33) combined conversational features, neuropsychological testing, and structural MRI to explore temporal features, and a linear mixed model was used to diagnose AD, which differs from our corpus. They also found that slow turn-taking and slow speech are two useful factors for the early detection of cognitive decline. Martinc et al. (34) also used a multimodal approach to detect AD on the ADReSS dataset, using an active data representation approach (13), combining linguistic, acoustic, and temporal features and obtaining an accuracy of 93.75%. Jonell et al. (35) recorded participants' language, speech, motor signs, pupil dilation, thermal emission, facial gestures, gaze, and heart rate variability of 25 patients with AD and found that multi-modality improved clinical discrimination. Recently, the transfer learning model has been widely used to diagnose AD. For example, Laguarta et al. (36) presented an approach with multiple biomarkers, including sentiment, lung and respiratory tract, and vocal cords. They used the transfer learning model to learn the features from audio datasets and obtained an accuracy of 93% on the ADReSS datasets. Zhu et al. (37) used the transfer learning on the BERT model to detect AD with speech and text, achieving an accuracy of 89.58%. They also found that the text model was more discriminative than the speech model. Overall, strong representation learning ability and discriminative classifiers, multimodal information, and transfer learning are all effective factors in the accurate diagnosis of AD and MCI.

Pre-trained models are considered important and effective nowadays because they attempt to learn the features and structure of the language from large datasets and regulate the model effectively to perform best on new datasets by only updating a few parameters. Accordingly, our model was highly trained with the best initial parameters. The best performance indicates that our model has learned useful features for classification, which not only reduces the need for expert-defined linguistic features but also makes it possible for accurate, complex, and comprehensive features to be extracted from the dataset. The advantage of sentence embedding is that it considers the entire transcript and does not have any out-of-context word embedding layer, which converts every word into a vector representation, considering its context. The ADReSS challenge also includes MMSE evaluation, a detailed interactive exam to evaluate cognitive skills, including memory, language, delayed recall, and visuospatial. However, whether our model is suitable for the evaluation of MMSE scores needs to be further verified. In addition, the transcripts were annotated in CHAT format (40), which is convenient for manual feature extraction. We performed the experiment with and without annotation and found that the performance did not differ. Using automatic speech recognition (ASR)-generated transcripts directly without the need for further annotation, our method has more advantages than the manual feature extraction method.

Many studies have demonstrated that manual features, combined with the deep learning model, can improve the performance of the model, and manual features also provide a better interpretation, which is important for clinical diagnosis. For example, Looze CD et al. (33) found that the temporal characteristics of speech may reflect underlying cognitive deficits. Nasreen et al. (32) used linguistic features, such as pauses, overlaps, and dysfluencies, to detect AD on the ADReSS dataset. They obtained 90% accuracy and demonstrated the importance of dysfluency and pauses in detecting AD. The champion of the ADReSS challenge (10) combined deep learning with pauses and obtained SOTA accuracy of 89.6%, proving that pauses are important for AD diagnosis. Sadeghian et al. (41) extracted acoustic features, including pauses more than 5 s in duration, and obtained the best accuracy of 95.8%. Features, such as pauses, are important features that deep learning cannot learn effectively (i.e., cannot give enough weight for pauses), so the combination of both can improve the performance of AD detection. In clinical medicine, patients with AD often pause and cannot continue treatment. This is not only a memory decline problem but may also be related to some language function obstacles caused by brain damage. A successful computer model can guide doctors to focus more on the early clinical symptoms of patients with AD, such as pauses and dysfluency. The largest limitation of our study is the difficulty to interpret the performance of a model with so many parameters (42). That is, our model cannot understand the reason for a wrong verdict, but we can identify the words that the network has paid more attention to in the case of a correct prediction. This function is particularly useful because such an interpretation can reveal the important linguistic attributes of patients with AD, which can help in speech therapy and communication with patients with AD.

The practice of pre-trained and fine-tuning paradigms has achieved excellent performance in many downstream tasks. In recent years, research in academia and industry has indicated that the pre-trained model is developing in a larger and deeper direction. However, there are still some problems that need to be solved in large models, such as the dataset quality, huge training energy consumption, carbon emission problems, and a lack of common sense and reasoning ability. These problems should be addressed in future studies.

In the future, we will focus on the following two directions for AD diagnosis.

Implicit sentiment analysis is an expression that does not contain any polarity markers but can still convey a clear human awareness sentiment polarity in the context; it exists widely in the recognition of aspect-based sentiments (43). For example, the comment “The waiter poured water on my clothes and walked away” contains no opinion words but can be interpreted as clearly negative toward “the waiter”; some sentences, such as “the service of the hotel is great,” “the food of the restaurant is delicious,” contain obvious sentimental words that neural network can give enough weights for the words of “great,” “delicious,” but the non-sentiment-related aspects of such words are often ignored by the model. The transcripts used for the AD diagnosis of spontaneous speech contain no polarity markers; however, most previous studies in this field generally pay little attention to implicit sentiment expressions. The study (44) used supervised contrastive learning to capture implicit sentiment using an advanced method. That is, the expressions with the same sentiment polarity were pulled together, and those with different sentiment orientations were pushed apart. In the future, we will focus on implicit sentiment analysis for AD diagnosis using a contrastive learning method.

One of the most popular language models is the multilingual one. With a proper multilingual model, the problem of lacking large datasets can be addressed by transferring the knowledge of AD prediction from another language in which a large dataset is available, which is similar to the approaches proposed by Fraser et al. (45). Only in this manner can the need for a target task be addressed for expert-defined linguistic features. In the future, we will commit to improving multilingual AD recognition using cross-lingual transfer learning, including the multilingual BERT and transformer models.

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found in the article/supplementary material.

Written informed consent was obtained from the individual(s) for the publication of any potentially identifiable images or data included in this article.

ZY designed the research. KL analyzed the data and interpreted the analysis. NL and YC wrote the main manuscript text and revised it carefully. All authors reviewed and approved the final manuscript.

This research was funded by Natural Science Foundation of Zhejiang Province (Grant Number LQ22H090002), the initial fee for introducing doctoral research of Natural Science Foundation of Zhejiang Provincial (Grant Number LGF20F020009) and the 4th Graduate Student innovation and Entrepreneurship Competition of Hangzhou Normal University.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.