In dementia, brain regions such as the hippocampus (involved in memory), language centers, frontal lobes, and communication pathways can be damaged, leading to language and memory problems. Many researchers have paid attention to speech data, and many methods were developed to distinguish dementia patients from healthy subjects [14,15,16,17,18,19]. These methods mainly used natural language processing techniques to evaluate the data. Different features were extracted and different types of machine learning models were proposed to perform classification on the text data, with labels indicating whether the text samples came from a dementia patient or a healthy subject. Weiner et al. [14] compared two feature extraction pipelines for dementia detection. The first pipeline involves manual transcriptions, while the second pipeline uses transcriptions created by automatic speech recognition (ASR). According to the study, the transcription quality of the ASR system is a dependable feature for detecting dementia on its own. Moreover, the features extracted from the automatic transcriptions exhibit comparable or slightly superior performance when compared to those derived from manual transcriptions. Mirheidari et al. [15] investigated how word vector representation models perform in detecting signs of dementia. Motivated by the fact that dementia patients suffer from impairment in accurately expressing something, they analyzed conversations designed to test examinees’ long-term and short-term memory. Three methods were proposed to show the potential of word vectors in a classification task. Their study concluded that it is possible to detect signs of dementia with a speech recognizer, even though the recognition result has a high error rate. Zhu et al. [16] found that transfer learning models utilizing text data outperform those using audio data, likely due to the high similarity between the pre-training text dataset and the Cookie-theft picture text dataset. Multi-modal transfer learning shows a slight improvement in accuracy, indicating that audio and text data provide limited complementary information. However, multi-task transfer learning results in limited improvements in classification and a negative impact on regression. They also identified that inconsistencies between AD/non-AD labels and Mini-Mental State Examination (MMSE) scores can limit the performance of multi-task learning. Mahajan et al. [17] re-implemented NLP methods that utilized Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) architectures and specific conversational features to diagnose AD. They examined why these models had lower accuracy on the ADReSS [20] dataset than the DementiaBank [21] dataset. Additionally, they created a deep learning-based solution using Recurrent Neural Networks (RNNs) that performed binary classification of AD from natural speech. Their method combined acoustic features using Speech-GRU, improving accuracy by 2% over the acoustic baseline. Furthermore, when enriched with some targeted features, their method outperformed acoustic baselines by 6.25%. They proposed a bi-modal approach to AD classification and discussed the potential benefits of their method. Shibata et al. [18] adapted and applied the idea density to the Japanese language. They defined several rules to count the idea density in Japanese speech and proposed a method to evaluate the idea density using machine translation. They proposed different pipelines for estimating the idea density, and compared the performance of different idea density calculation methods. They concluded that these findings prove that it is feasible to apply the method in the area of dementia detection. Santander-Cruz et al. [19] proposed a method in which they extracted 17 kinds of features, which represented syntactic, semantic, lexical, and demographic information of the speech samples collected in the DementiaBank Pitt Corpus dataset. The relevance of the 17 features was quantified by calculating a mutual information score that can represent the dependency between the extracted features and the MMSE score. They concluded that their methodology is superior to the methods based on syntactic information or the BERT method with only linguistic features. Farzana et al. [22] evaluated how removing the disfluency in the annotated result of an automatic speech recognition system affects the performance of dementia detection. They used an existing tool that detects and tags disfluency in transcriptions of speech. Through a set of experiments, they found that removing the disfluency has a bad influence on dementia detection performance, which reduces the detection accuracy. Ilias et al. [23] proposed a novel method that combined two modalities such as speech and transcripts into one model based on the vision transformer. They used the gated multimodal unit to control how much influence each modality had on the final classification. They also used crossmodal attention that learned the relationships between modalities in an effective way. They tested their work on the ADReSS Challenge dataset and proved their model’s superiority over existing methods.

The above-mentioned most recent studies concentrate on the use of speech and text data to detect dementia. However, in some past studies [24,25] which analyzed the facial expressions of dementia patients, some conclusions mentioned that some deviations happen among dementia patients. The different displays of facial expressions have the potential to detect dementia. Asplund et al. [24] showed that patients with some types of dementia tend to show fewer facial expressions. They analyzed the ability to produce expressions using the Facial Action Coding System (FACS) under pleasant and unpleasant stimulus conditions, while, in their later study [25], they compared two methods developed for interpreting the facial expressions of demented people. In this research, disagreement between the two methods was reported, and demented people had less clarity or a lower amount of facial cues. Based on the findings of this research, It can be noticed that the facial expressions of demented people deviate, which may function as indicators in dementia detection.

In recent years, some works [12,13,26] collected facial expression data from dementia patients and explored the potential of using facial expressions to detect dementia. Tanaka et al. [12] collected human-agent interaction data of spoken dialogues from 24 participants. They extracted facial features from these data and used L1-regularized logistic regression to classify dementia patients and healthy subjects. Their research identified several features, including Action Units (AUs), eye gaze, and lip activity, as contributing to the classification. They also identified the importance of each feature in their L1 regularized logistic regression. The majority of the important features are AU-derived features. However, their methods do not apply to natural free conversation, nor did they evaluate their methods on a large dataset. Umeda et al. [13] examined whether artificial intelligence could distinguish between the faces of people with and without cognitive impairment. They achieved a classification accuracy of 92.56%. Their study showed that it is possible for deep learning to distinguish the faces of dementia patients from those of people without dementia. Although their study achieved high accuracy, they used facial images for the study, and institutional bias was reported [13]. Institutional bias can mislead the machine learning model so that the model learns environmental information instead of the dementia indicators.

Some papers have analyzed the expressions of dementia patients under controlled conditions. However, they did not apply them to the detection of dementia: Liu et al. [26] investigated how different sound interventions affect the facial expressions of older people with dementia. They showed the participants different sound pieces and analyzed their emotions. Jiang et al. [27] used a computer vision-based deep learning model to analyze facial emotions expressed during a passive viewing memory test in people with cognitive impairment caused by AD and other causes. Their results showed evidence that facial emotions are different in patients with cognitive impairment. Specifically, patients with cognitive impairment showed more negative emotional expressions and more facial expressiveness than healthy controls. They concluded that facial emotions could be a practical tool for screening patients with cognitive impairment.