We used conversational data obtained during phone calls with a regular monitoring service for seniors provided by Cocolomi Co, Ltd (Figure 1A). In this service, communicators call and talk with older individuals once or twice a week, typically for 10 to 20 minutes of free conversation. All communicators received training and communicated with participants in accordance with the company guidelines for encouraging participants to speak. In this service, one communicator is typically assigned to a participant, and that participant talks with the same communicator. They manually transcribe the conversations in a spoken-word format by omitting incomplete words and fillers and forward the texts to family members such as their children. They used either home phones or smartphones. All conversations were conducted in Japanese. We analyzed these transcribed text data. All participants agreed to having their conversational data used for research purposes. This study was conducted under the approval of the ethical committee of Shizuoka University and performed in accordance with the Ethical Guidelines for Medical and Health Research Involving Human Subjects.

The data were obtained from 15 Japanese people (12 females and 3 males aged 61 to 91 years; mean 76.8 [SD 9.4] years). Of these, 2 females had received the diagnosis of AD. The morbidity information for other diseases was not available. The communicators knew the diagnosis status of the participants, but they communicated with them in accordance with the company guidelines and did not change conversation methods on the basis of the diagnosis status. The follow-up period ranged from 1 to 33 months (mean period 16.1 months). We analyzed data from 1032 phone calls in total, and the number of phone calls for each participant ranged from 4 to 226 (mean 68.8 phone calls). The average duration of a single phone call for each participant ranged from 6.8 to 22.2 minutes (mean time between participants 12.1 [SD 4.1] minutes). The total call time of our dataset was around 221 hours. Transcribed text data consists of only text spoken by participants, not communicators. The text data contained 1098 characters on average, and we analyzed 1,132,935 characters in total. Table 1 provides the overview of our dataset.

From each item of phone call data, we extracted linguistic features related to word and topic repetition within single conversations, as well as the standard linguistic features used in previous studies to build models for AD screening from connected speech data during neuropsychological tests (for the list of features, please see the Standard Linguistic Features section) [10,14-34]. In addition, we extracted features related to word and topic repetition across different conversations from the text data of two phone calls separated by t days and n phone calls interval, respectively (Figure 1B). Specifically, we used all possible pairs of two phone call data items separated by t days (T–M<t≤T+M) for specific days interval and n phone calls (N–M<n≤N+M) for specific phone calls interval. Optimal parameters for T (4,5,…,13) and N (3,4,…,12) were selected by performing a grid search for each feature. In this study, we used M=3, T=10, and N=7 for word repetition, and T=7 and N=9 for topic repetition.

For preprocessing, we used the Japanese morphological analyzer MeCab [42] to perform word segmentation, part-of-speech tagging, and word lemmatization on the transcribed texts. Words tagged as numerals or symbols were excluded from the analysis. Predefined stop words were also eliminated.

We measured word repetition by using the feature focusing on the number of distinct words being used in a target document. Specifically, the feature was calculated by the inverse number of Honoré’s statistic (HS) [43], defined as HS=100logV/(1–V_uni/U), where V is the total number of words, U is the total number of distinct word types, and V_uni is the number of total distinct word types used only once. For the feature related to word repetition across two text data of different conversations, we combined the data of both texts and then extracted the feature from it.

For the features related to topic repetition, we used a biterm topic model (BTM) [44]. While conventional topic models such as latent Dirichlet allocation have typically been used for estimating the latent topics of conversations [45], BTM was designed to extract latent topics from a limited amount of context, such as that found in tweets and news headlines [44]. Accordingly, we used BTM for estimating the topic in sentences and then extracted features related to topic repetition within conversations. Specifically, we first divided a text document into sentences by using punctuation marks. We then applied a BTM to a set of all the sentences, extracted word probability vectors for each K topic and obtained topic proportions allocated to each l sentence. For example, when a text document consisted of L sentences, we obtained L–l+1 vectors for topic proportions. We calculated pair-wise similarities with Euclidean distance and used the inverse number of the mean value as topic similarity features within single conversations. For the topic similarity features across two different items of text data D_i and D_j, we calculated pair-wise similarities between topic proportions in D_i and D_j and used the inverse number of the mean value. The parameters K=200 and l=3 were used.

We briefly describe how BTM works. In contrast to conventional topic models such as latent Dirichlet allocation [45], BTM directly models the generation of word co-occurrence patterns in the whole corpus to estimate topics from a small amount of context in a document [44]. The key concept is that if two words co-occur more frequently, they are more likely to belong to the same topic. The term biterm denotes an unordered word pair occurring in a short text. For example, a document with three distinct words generates three biterms:

In general, a document containing n distinct words generates nC₂ biterms. The biterms are typically extracted from each sentence, and by combining them we get a corpus B for the target document as a set consisting of ‖B‖ biterms. With the BTM, it is assumed that each topic is a multinomial distribution over the biterm (ϕ), and the whole corpus is a mixture of these topics (θ). The process of extracting biterms can be conducted as follows:

Following the above procedure, the probability of biterm b_i=(w_i,₁, w_i,₂) conditioned on the model parameters θ and α can be written as seen in Figure 2. For more details, we refer the reader to the original paper [44].

On the basis of related work on connected speech data during neuropsychological tests, we extracted 29 linguistic features categorized into 4 types of linguistic features that would be useful for inferring AD: parts of speech, vocabulary richness, syntactic complexity, and perseveration [10,14-34]. Here, neuropsychological tests in previous studies include a picture description task using a picture of the Cookie Theft, comprehensive aphasia test, and the Wechsler Logical Memory tests [46-48].

The first feature set, related to parts of speech, consists of 14 features: frequency and the ratio of part-of-speech tags. We extracted the part-of-speech information by using Mecab [42] and then computed the frequency of occurrence of different parts of speech (nouns, verbs, adjectives, auxiliary verbs, and conjunctions). We also computed ratios, namely, each part of speech normalized by the total number of word tokens in the document. Analyzed part-of-speech tags included nouns, verbs, adjectives, pronouns, adverbs, auxiliary verbs, and conjunctions. We also calculated the ratio of noun to verb and the ratio of pronoun to noun.

The second feature set, related to vocabulary richness, consists of 3 features: type-token ratio (TTR), Brunét’s index (BI), and HS [11,43]. This feature set measures lexical diversity, which tends to decrease in AD cases [11,13]. TTR compares the total distinct word types (U) to the total word count (V) as TTR=U/V. Using the same U and V, BI is defined as BI=V^U–0.165. Unlike other measures related to vocabulary richness, for this measure, the lexical richness becomes greater as BI becomes smaller. HS gives particular importance to unique vocabulary items used only once, also known as hapax legomena V_uni. HS is defined as HS=100logV/(1–V_uni/U).

The third feature set, related to syntactic complexity, consists of 7 features related to length metrics as well as dependency relations. For features related to length metrics, we calculated 4 measures consisting of mean length of sentences, total number of sentences in a document, total number of words in a document, and total number of characters in a document. For features related to dependency relations, we calculated total number of dependencies in a document, the average dependencies per sentence, and total dependency distance in a document. As for the dependency structures, we used CaboCha [49]. Dependency distance was calculated by the sum of all the linear distances between 2 syntactically related words in a sentence. This dependency distance has been considered an important index of memory burden and an indicator of syntactic difficulty [50].

The final feature set, related to perseveration, consists of 5 features. First, sentences were converted into term frequency–inverse document frequency (TF-IDF) vectors by using a bag-of-words model [51]. TF-IDF is a numerical statistic intended to reflect how important a word is to a document. By using TF-IDF vectors, we then calculated the cosine similarity between sentences [51]. We then calculated the proportion of sentence pairs equal to 0 and below 2 thresholds (0.3, 0.5) as well as the minimum and average cosine distances across all pairs of sentences.

All statistical analyses were done in the R environment (R Foundation for Statistical Computing). A 2-tailed Student t test was applied for each feature to determine significant differences of enrichment between the two populations, and the resulting P values were adjusted with Bonferroni multiple testing correction. Adjusted P values below .01 were considered significantly different.

In response to the increasing demand for detecting dementia in everyday situations, we aimed to quantify language dysfunctions observed in AD from daily conversational data when the target is not performing neuropsychological or cognitive tasks. Although previous studies have succeeded in quantifying many aspects of language dysfunction, including grammatical and informational content as well as speech characteristics resulting from AD, they mainly investigated connected speech data collected while targets were engaged in neuropsychological and cognitive tasks; how language dysfunction in daily conversations can be quantified has not been sufficiently investigated. In this study, we focused on atypical repetition of words and topics reported by previous observational and descriptive studies as one of the prominent characteristics in everyday conversations of AD patients. Using conversational data of seniors with and without AD obtained from longitudinal follow-up in a regular monitoring service (from n=15 individuals including 2 AD patients at an average follow-up period of 16.1 months; 1032 conversational data items obtained during phone calls and approximately 221 person-hours), we investigated whether and how linguistic features related to word and topic repetition can be used for differentiating AD patients.

The results indicated that atypical repetition across different conversations could have a better discriminative power for AD compared to that within single conversations. We assume this atypical repetition across different conversations results from memory impairment that prevents speakers from remembering recent conversations. If so, the results suggest that the difference of memory performance between AD and healthy older adults might be larger after the lapse of some time than immediately, and linguistic features relating to atypical repetition could capture such difference even in daily conversations. In fact, deficits in episodic memory are one of the earliest detectable cognitive impairments in AD [53], and some experimental studies on episodic memory tasks such as word-list learning have reported that the forgetting rate passing from immediate to delayed recall in AD was significantly larger than that in age-matched controls [54]. These experimental studies seem to support our results. In addition, our results suggest that interval days between two conversations might be a more important parameter than the number of phone calls in terms of a discriminating power. These results might also be reasonable if the linguistic features relating to repetition can capture memory impairments resulting from pathological changes underlying AD. Furthermore, recent studies have reported accelerated long-term forgetting even in presymptomatic AD, which has garnered increased attention to detecting presymptomatic changes in AD [55]. From this perspective, our approach, which focuses on atypical repetition in daily conversations on different days, holds promise for detecting early signs of AD in everyday situations.

On the basis of comparisons with the standard linguistic features used in previous studies on connected speech data during neuropsychological tests, we found that features relating to atypical repetition in two conversations on different days had better AUC-ROC scores for differentiating AD than these linguistic features on speech data of daily conversations. The results suggest that repetition features across daily conversations on different days might be better indicators for detecting AD in everyday situations. We also compared the results of statistical analysis for each linguistic feature between previous studies on connected speech data during neuropsychological tests and our studies on daily conversations and found that about 80% of features (19 out of 24) showed a statistically consistent tendency. These results indicate that these linguistic features could be useful for inferring AD from speech data not only during neuropsychological tests but also during daily conversations. Features showing a statistically different tendency included those related to part-of-speech distributions and perseveration. In terms of the features of part-of-speech distribution, adjective ratio and noun to verb ratio showed significant difference between AD and controls in our study, while they were reported to have no significant difference in previous studies on connected speech data during neuropsychological tests [15,18,22,24]. In contrast, previous studies on speech data during semistructured interviews have reported that these two features were significant different [11,56]. Thus, language dysfunction in AD measured by these two features might be discriminative, especially in conversations in which the contents are less predetermined. As for the features related to perseveration, we found no significant difference in our dataset, while previous work reported the statistical significance of these features [22]. For example, one study used these features for measuring semantic similarity using a bag-of-words model in speech data during a picture description task and reported a significant difference between AD and controls [22,57]. Considering our results on the statistical analysis for these features, when we analyze unstructured and free conversational data such as daily conversations, language dysfunction in AD related to perseveration might need to be measured by combining a model for estimating the semantic topics (eg, BTM). In fact, the semantic similarity when using the BTM revealed a significant difference between AD and controls in our dataset.

In this study, we analyzed daily conversational data and showed that linguistic features, especially those related to repetition, can potentially be used for automatically detecting AD in a longitudinal and passive manner. Indeed, speech data collected in daily lives has gained increasing interest for clinical applications due to the improvement in audio quality recorded by portable devices and the expansion of voice-based interaction systems such as smartphones and smart speakers [4,58]. In particular, several studies that administrated cognitive and neuropsychological assessments over the phone have reported that speech features could be reliably extracted from phone recordings and used to build models for AD screening [59]. However, there are still relatively few studies on changes in speech features in daily conversations resulting from AD. Being capable of inferring AD from daily conversations would help with timely detection by frequent assessments with relatively short intervals, which might be difficult to do when using neuropsychological tests due to learning effects. From this perspective, we believe that the results of our study will help promote future efforts toward early detection of AD in everyday situations.

Our work has several limitations. First, the number of participants was small, especially for AD patients, although the follow-up periods and number of conversations for each participant were relatively long and large. The results of a post hoc power analysis revealed that we would need at least two additional AD patients to obtain more than 0.80 of the power. A second limitation is the lack of information about AD severity and stage. We analyzed conversational data collected from the users of a regular monitoring service, and as such, were not able to obtain information related to clinical assessments, such as Mini-Mental State Exam scores and biomarkers (aside from AD diagnosis provided by medical doctors). Third, the age and gender of the control and AD groups were not matched. These three limitations stem from the fact that our dataset was collected from real users of an actual monitoring service. To clarify the relevance of our results obtained from the datasets with these limitations, we comprehensively compared the statistical results of each of the 24 linguistic features between our study and more than 20 previous studies and confirmed that our dataset showed a statistically consistent tendency with the previous studies. In addition to this perspective, because our dataset represents invaluable data collected from a real service and consists of 1032 conversational data items at an average follow-up period of 16.1 months, we believe our results can provide useful information for future studies toward the early detection of AD in the real world. The fourth limitation is the use of transcribed conversation data after omitting information related to incomplete words and fillers. While previous studies on connected speech during neuropsychological tests reported that these features did not significantly differ between controls and AD patients [40], we feel that further investigation is required, especially on speech data in daily conversations. The fifth limitation was the specific type of conversational data used. In this study, we analyzed free conversation between older adults and communicators mainly about the lives of the older individuals, but we need further research to confirm our results and extend the scope into various types of daily conversation, such as phone call conversations with their friends and face-to-face family conversations.

In summary, we investigated the daily conversational data of seniors with and without AD obtained from longitudinal follow-up in a regular monitoring service. We demonstrated that, while the linguistic features used in previous studies on connected speech data during neuropsychological tests can be used for detecting AD from daily conversations, the novel features related to repetition across conversations on different days could be a better indicator. We believe that our results can help to promote future efforts toward early detection of AD in everyday situations by taking advantage of speech data that can be collected in a passive, longitudinal manner.