Sage Journals: Discover world-class research

Abstract

Introduction

As the incidence and prevalence of dementia continue to rise, there is a critical need for more cost- and time-efficient diagnostic tools. Analysis of speech prosody has emerged as a promising noninvasive biomarker, potentially offering a more accessible approach to dementia identification. However, the absence of a longitudinal analysis of thematic evolution within the extensive literature in this domain has resulted in a notable knowledge gap.

Methods

We conducted a text mining analysis of publications from the past 30 years to identify key research trends, thematic patterns, and associated topics.

Results

Our analysis yielded three major findings: a marked acceleration in research activity since 2020, a convergence of clinical needs with technological advancements, and the inherently interdisciplinary nature of this field.

Discussion

These findings not only underscore the dynamic evolution of dementia research but also highlight the potential of speech prosody analysis as a viable, noninvasive diagnostic tool. Future research integrating multidisciplinary approaches and evaluating diagnostic values of speech prosody is warranted.

Keywords

dementia speech biomarker text mining digital biomarker diagnosis

Introduction

Dementia is a progressive neurological condition characterized by a decline in multiple cognitive domains, including memory, language, problem-solving, and executive function. The cognitive decline interferes with daily life and independence in people with dementia. According to the Alzheimer's Society,¹ dementia of Alzheimer's type (DAT) is the most prevalent form of dementia, followed by other types such as vascular dementia (VaD), dementia with Lewy bodies, and frontotemporal dementia (FTD). While cognitive decline in dementia is well documented, less attention has been given to communication deficits, specifically changes in prosody—defined as the rhythm, pitch, and intonation patterns of speech. Prosody, an element of speech that can be measured acoustically, plays an important role in everyday communication by delivering the speaker's intent beyond the literal meaning. This speech element consists of two major categories, linguistic and affective. While linguistic prosody provides cues regarding syntax and pragmatics, affective prosody plays an important role in everyday communication by delivering the speaker's emotion.

Prosodic impairments are common among individuals with dementia, and understanding these changes has become an emerging area of interest. Research suggests that affected individuals may exhibit reduced pitch variation, slower speech rate, and altered rhythm, all of which affect their ability to manipulate prosodic features for effective communication.^2,3 For example, reduced pitch range can lead to flat or monotone speech, impacting listeners’ perceptions of the speaker's affective state.³ These prosodic changes may differ across dementia types; for instance, individuals with VaD may speak more softly than those with DAT.³ Identifying these distinctive patterns could not only aid in distinguishing between dementia subtypes but also support early diagnosis and monitor cognitive decline over time, as prosodic impairment often precedes more pronounced cognitive symptoms.^4,5 With regard, research has shown a growing interest in exploring the link between prosodic alterations and dementia, with findings indicating that speech prosody analysis can offer unique insights into neurocognitive disorders as a promising tool for dementia diagnosis.^2,3 However, there is a lack of a comprehensive review demonstrating how this area of research has evolved.

Understanding how a research area has evolved is crucial because it provides context, highlights trends, and identifies gaps in existing knowledge.⁶ This understanding allows researchers and practitioners to see how theories, methodologies, and findings have developed over time, which can guide future research, avoid redundancy, provide a historical perspective, enhance practical application, and validate or challenge existing assumptions.^6,7 Therefore, grasping the evolution of research on speech prosody in dementia diagnosis is crucial for advancing this complex field. Early studies established the foundational idea that changes in speech patterns, specifically prosody, may be a marker of cognitive decline.⁸ However, advances in technology and methodologies have enabled scientists to pinpoint specific prosodic elements, such as frequency, amplitude, and timing, that may differentiate between cognitive states with greater precision. Despite these advancements, the field lacks a comprehensive review that connects early insights with contemporary findings and methodologies. This gap in synthesis leaves today's researchers with a fragmented understanding, potentially leading to redundant studies or efforts that do not capitalize on recent technological innovations.⁹ A longitudinal analysis of the literature on speech biomarkers in dementia can offer valuable insights into how the field has evolved over time and inform future directions.

The overarching objective of this study was to identify longitudinal research trends and contextual shifts in the field of speech biomarkers for dementia diagnosis over the past three decades. To achieve this, we employed text mining techniques to analyze a broad corpus of publications. Our primary goal was to uncover thematic evolution across time, exploring broad research trends. As a secondary aim, we exampled the methodological approaches used in the existing literature. Within this broader landscape, these changes in prosody may thus hold potential as early indicators of cognitive decline and specific types of dementia, enhancing both clinical diagnosis and ongoing assessment of disease progression.

Methods

Search strategy

We performed a comprehensive, systematic search across five key databases: PubMed, Web of Science, CINAHL, Scopus, and Linguistics & Language Behaviors Abstract. To ensure consistency in search terms and enhance the sensitivity of the search strategy, both Medical Subject Headings (MeSH) and predefined subject headings were employed. A librarian recommended the combined use of MeSH and predefined subject headings. These terms were applied either independently or in combination, incorporating dementia- and prosody-related terms. The searches were restricted to peer-reviewed publications over the past 30 years written in English.

The search terms related to dementia were dementia, FTD, VaD, Alzheimer's disease, Lewy body disease, and primary progressive aphasia. These terms were selected based on dementia's definition and the widespread prevalence of its pathology. Dementia, a broad term rather than a single disease, refers to “the loss of cognitive functioning” that limits an individual's ability to perform everyday tasks.¹⁰ This classification groups multiple pathologies by shared characteristics, such as common causes (e.g., protein build-up) or specific brain regions affected.¹¹

The prosody-related search terms included speech production measurement, speech acoustics, prosody, pitch, intonation, speech stress, and speech acoustics. These broader acoustic terms were included to maximize sensitivity in capturing studies that may investigate prosodic aspects of speech, even if not labeled explicitly as “Prosody.” During the screening, studies were reviewed to ensure relevance to prosodic features such as intonation, rhythm, and stress, and studies focusing exclusively on spectral or voice quality measures (e.g., jitter and shimmer) without a clear connection to prosodic analysis were excluded. Prosody involves a range of acoustic features that go beyond individual sounds, including intonation patterns (or pitch contours) shaped by the fundamental frequency (F0) of the voice; timing cues that help listeners recognize boundaries between words, phrases, and full statements; and the stress placed on words and sentences, a nuanced quality influenced by volume, pitch, and length.¹² As speakers incorporate prosodic elements while forming and expressing each thought, these features offer intriguing possibilities for dementia assessment, as they may reflect underlying cognitive processes.

The search strategy was reviewed by an independent librarian, as recommended by Crowther et al.¹³ A dual independent review was applied to the search strategy and to the processes of identification and assessment of studies to reduce random errors and bias per recommendation by Eden et al.¹⁴

Identification of studies

The search results were downloaded to Microsoft Excel version 16.91 and went through a qualification screening process, including duplicate removal and title and abstract screening. Duplicates were removed using the Microsoft Excel conditional formatting function, followed by manual checks. Titles and abstracts were screened for relevancy by two independent reviewers. The two researchers’ independent work reached an initial agreement of 97.8%, and the discrepancy was resolved via discussions until the agreement reached 100%. The inclusion and exclusion criteria used for title and abstract screening are described in Table 1.

Table 1.

Inclusion and exclusion criteria.

Criteria	Inclusion	Exclusion
Types of studies	Studies published between 1992 and 2023 Peer-reviewed Original research article Written in English	Conference proceedings Case reports/descriptions Book chapters
Study participants	Any type of dementia	Studies not clearly including people with a dementia diagnosis Studies exploring dementia caregivers’ prosody
Target outcomes	Expression and comprehension of speech prosody	Studies not employing acoustic analysis Studies focusing on other aspects of prosody (e.g., emotional processing) than the acoustic production of speech prosody
Outcome measures	Any outcome measured by any type of measurement at any time

Data analyses

We used the author keywords to identify major themes and grouped individual keywords into the identified themes. The author keywords serve as representations of an article's content and highlight the central themes and research focus, indicating the fields, subfields, topics, and approaches pursued by the researchers.^15–17 The keywords were grouped manually by the two independent researchers from the collected literature, and the grouping was discussed until they reached 100% agreement.

We also reviewed the abstracts of the collected literature to further understand the overview of the research trends in the field, focusing on topics and approaches. In this phase, text mining was utilized to support the literature review process. Text mining is increasingly recognized as a beneficial tool for conducting literature reviews across various disciplines due to its ability to rapidly uncover patterns and insights that might otherwise remain obscured in large bodies of literature.^18,19 A key advantage of this computer-assisted method lies in the capacity to extract and synthesize knowledge from texts using a range of metrics (e.g., co-occurrence networks, term frequency-inverse document frequency). These techniques facilitate the identification of trends, relational dynamics, and overarching themes that may not be readily apparent through traditional review approaches.¹⁸ By integrating these computational techniques, the review process gains a more nuanced understanding of the research landscape, offering critical insights that can inform future research directions.^20,21

To leverage these strengths, text-mining techniques were used in this study by utilizing KHCoder 3^22,23, a text-mining software for computational linguistics. In this study, term frequency (tf), document frequency (df), and co-occurrence network of words techniques were employed to understand major topics extracted from the abstract by year and associations among the extracted terms. Tf measures how often a term appears in a corpus, whereas df of a term refers to the number of documents in which the term appears in the corpus.²⁴ Both tf and df help identify the importance of a term in the context of the entire collected literature. Co-occurrence analysis is to explore the relationships between terms that frequently appear together within the same document (i.e., an abstract in the current study).²⁵ The degree of association between terms in the network was calculated using the Jaccard coefficient (J_c)²⁶ which varies between 0 and 1 based on how frequently the terms appear in the same abstract.

Results

Search outcomes

Figure 1 illustrates the search flow and analytic pipeline. The database search yielded 4522 articles, and 54 of them were removed due to duplication. For the remaining articles, titles and abstracts were screened, and the final outcomes contained 110 relevant articles.

Figure 1.

Search flow and analytic pipeline.

Temporal research trends

Figure 2 illustrates the progression of publication on speech prosody in relation to dementia diagnosis. Over the past three decades, the total number of publications in this field gradually increased from 1997 to around 2015. However, until 2020, the number of peer-reviewed articles remained relatively low, typically ranging from 1 to 9 per year. A marked shift occurred starting in 2020, with the number of publications increasing rapidly from 11 to 34 and reaching a peak in 2023. This remarkable growth was most evident in articles indexed in Scopus and Web of Science.

Figure 2.

Distribution of publications by year.

Key research topic areas and methodological approaches

A total of 461 unique author keywords were identified from the collected literature, while 26 articles did not include any author keywords. Based on the keywords, four major thematic clusters were determined: Diseases, Study Subjects, Methodology, and Study Design. Keywords regarding Diseases and Methodology emerged as relatively more prominent compared to the others, Study Subjects and Design, exhibiting a higher frequency of dominant keywords. Table 2 illustrates the four themes and the five most frequently used keywords in each theme.

Table 2.

Top 5 most frequently used author keywords by themes.

Rank	Themes and term frequencies (tf)
Rank	Disease	tf	Study subjects	tf	Methodology	Tf	Study design	tf
1	Alzheimer's disease	34	Human	12	Speech analysis	38	Comparative study	6
2	Primary progressive aphasia	18	Aged	11	Machine learning	17	Evaluation research	4
3	Dementia	14	Female	8	Acoustic analysis	16	Multicenter studies	3
4	Parkinson's disease dementia	13	Male	7	Statistical analysis	14	Clinical trial	1
5	Mild cognitive impairment	9	Middle age	5	Prosody analysis	12	Explorative research	1

To uncover distinct temporal patterns in the research priorities within the field, a crosstab analysis was carried out, focusing on the types of dementia studied, the methodologies employed, and the purposes of the research. Thematic trends identified from the author keywords were mapped onto the abstract to reveal the shifts in research priorities over time.

We discovered a notable emphasis on speech and acoustic analysis in the early 2000s, with terms such as prosody, speech analysis, and acoustic analysis appearing frequently. This suggests that researchers during this period were actively exploring speech-based markers as a noninvasive means for screening and early detection of dementia, especially for Alzheimer's disease, which consistently remained the most studied condition across all years. Around 2016, a methodological shift became apparent, marked by the growing appearance of terms such as machine learning, neural networks, and later deep learning. This finding indicates the increasing adoption of computational methods, particularly machine learning–driven approaches, culminating in a significant peak of machine learning–related research around 2020. These newer techniques were often applied in conjunction with natural language processing, highlighting a movement toward leveraging computational methods for analyzing unstructured data such as patient speech or clinical notes. In terms of dementia subtypes, while Alzheimer's disease has remained central, there was also a surge of interest in conditions such as FTD, Parkinson's disease, and primary progressive aphasia during the mid-to-late 2000s. This broader focus indicates a growing recognition of the heterogeneity of dementia and the need for diagnostic tools tailored to its various forms (Figure 3).

Figure 3.

A bubble plot by year and major.

Within the theme of disease, Alzheimer's disease (tf = 34) emerged as the predominant focus, with a term frequency more than twice that of any other disease. Four other most frequently used keywords following Alzheimer's disease were primary progressive aphasia (tf = 18), dementia as a general term (tf = 14), Parkinson's disease (tf = 13), and mild cognitive impairment (tf = 9). Terms within the theme of study subjects were generally generic and infrequent with the highest-frequency term, Human, appearing in only 12 articles. In the Methodology, speech analysis emerged as the most frequently mentioned term, appearing nearly twice as often as the second-most frequent term, machine learning (tf = 17).

To determine which methodological approaches are interconnected with dementia research, we computed similarity scores and visualized the associations of terms from the abstract using the co-occurrence technique (Figure 4). This analysis is to clarify the relationships between the methodologies employed, the identified types of dementia, and their purposes (e.g., diagnosis, screening, or detection). The notable correlations between the term “Alzheimer's disease” and “speech” (J_c = 0.27), as well as the term “speech analysis” (J_c = 0.215), “detection” (J_c = 0.250), and “diagnosis” (J_c = 0.298) showed that speech analysis and speech features were used to help diagnose or detect Alzheimer's disease rather than for screening (J_c = 0.155). Furthermore, the relatively closer link between the term “machine learning” and “Alzheimer's disease” (J_c = 0.20) also indicates machine learning techniques were used in relation to Alzheimer's disease. It is noteworthy that the term “Amyotrophic Lateral Sclerosis”, although not included in our original search terms, emerged with relatively strong co-occurrence values with “speech” (J_c = 0.276), detection (J_c = 0.232), and diagnosis (J_c = 0.25), but showed weaker association with screening (J_c = 0.083).

Figure 4.

A modularity network of terms regarding dementia types, methodologies, and research purposes extracted from the abstract (N 18, E 50, D. 327).

A similar pattern was observed among the terms “primary progressive aphasia”, “speech” (J_c = 0.231), and diagnose (J_c = 0.21), indicating speech features were also used to diagnose PPA. However, unlike Alzheimer's disease and ALS, the term “primary progressive aphasia” also showed relatively stronger associations with acoustic (J_c = 0.244) while showing a relatively weaker association with “detection” (J_c = 0.068).

It was also found that machine learning techniques were relatively more often utilized in detection (J_c = 0.25) and screening (J_c = 0.276) rather than diagnosis (J_c = 0.113). Among all dementia types, Alzheimer's disease–related studies more actively adopted machine learning techniques (J_c = 0.20) than other types of dementia. The term “machine learning” was found to be associated with traditional methods such as “speech” (J_c = 0.130) and “speech analysis” (J_c = 0.179), “acoustic” (J_c = 0.107) and “acoustic analysis” (J_c = 0.139). Although these associations were relatively moderate, they indicated a growing interest in applying machine learning techniques to analyze speech and acoustic features in the context of dementia. The associations between the term “machine learning” and the terms “screening” (J_c = 0.276) “detection” (J_c = 0.250), and “diagnosis” (J_c = 0.113) demonstrated that this computer-assisted method is more commonly adopted in the screening and detection of dementia than in its diagnosis (see Table 3 for details).

Table 3.

Similarity scores across analysis techniques and dementia types.

	AD	Acoustic	AA	ALS	DL	Detection	Diagnosis	FTD	ML	MCI	NLP	NN	PD	PPA	Prosodic	Prosody	Screening	Speech	SA
AD	-
Acoustic	0.150	-
AA	0.145	0 . 307	-
ALS	0.289	0.175	0.065	-
DL	0.019	0.025	0.000	0.048	-
Detection	0.250	0.206	0.182	0.232	0.089	-
Diagnosis	0.298	0.308	0.167	0.250	0.030	0.355	-
FTD	0.083	0.082	0.000	0.134	0.000	0.073	0.114	-
ML	0.200	0.107	0.139	0.069	0.136	0.250	0.113	0.000	-
MCI	0.143	0.096	0.054	0.071	0.000	0.208	0.069	0.000	0.097	-
NLP	0.000	0.013	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.062	-
NN	0.019	0.025	0.036	0.048	0.300	0.067	0.047	0.000	0.043	0.000	0.000	-
PD	0.116	0.143	0.130	0.189	0.029	0.164	0.154	0.000	0.022	0.047	0.000	0.029	-
PPA	0.090	0.244	0.053	0.185	0.023	0.068	0.210	0.200	0.019	0.000	0.000	0.024	0.048	-
Prosodic	0.125	0.260	0.104	0.158	0.018	0.120	0.179	0.115	0.078	0.047	0.000	0.018	0.111	0.158	-
Prosody	0.089	0.189	0.070	0.169	0.000	0.096	0.125	0.146	0.037	0.038	0.000	0.000	0.081	0.154	0.745	-
Screening	0.155	0.106	0.105	0.083	0.040	0.271	0.176	0.000	0.276	0.250	0.000	0.000	0.067	0.000	0.094	0.056	-
Speech	0.270	0.545	0.184	0.276	0.035	0.262	0.373	0.122	0.130	0.133	0.009	0.035	0.195	0.231	0.315	0.220	0.129	-
SA	0.215	0.352	0 .109	0.189	0.038	0.268	0.267	0.140	0.179	0.143	0.021	0.019	0.069	0.076	0.269	0.194	0.218	0.391	-

Note. AA: acoustic analysis; AD: Alzheimer's disease; ASL: amyotrophic lateral sclerosis; DL: deep learning; FTD: frontotemporal dementia; MCI: mild cognitive impairment; ML: machine learning; NLP: natural language processing; NN: neural network; PPA: primary progressive aphasia; PD: Parkinson's dementia; SA: speech analysis. Values greater than 0.1 are shown in bold.

Discussion

The overarching aim of the current research was to explore the emersion of speech prosody as a potential tool for dementia diagnosis. To achieve this aim, we identified research trends and themes over the past 30 years in this area and explored associated research topics using text mining. Our study revealed several significant patterns in the research landscape of speech prosody as a potential diagnostic tool for dementia.

First, our analysis of research trends demonstrated a marked acceleration in research activity since 2020 in studies investigating speech biomarkers for dementia. This surge aligns with two concurrent developments in the field: the growing emphasis on noninvasive biomarkers and the adaptation to remote healthcare delivery during the COVID-19 pandemic. An interest in noninvasive and cost-efficient dementia diagnostic methods as an alternative to expensive and invasive tests such as neuroimaging or cerebrospinal fluid analysis has been growing. Traditional dementia diagnosis, relying on expensive neuroimaging and invasive biomarker tests, makes it difficult for older adults to use it as routine screening.²⁷ However, speech analysis may provide a noninvasive and cost-efficient alternative that utilizes everyday communication activity to identify dementia.^3,8 More specifically, studies on speech prosody have shown that modulations in prosodic features such as rhythm, pitch, and stress can serve as indicators of dementia.⁸ The timing of this increase is particularly noteworthy, as it coincides with the healthcare system's urgent need for remote diagnostic tools, suggesting that practical necessity may have accelerated research in this direction. Due to restricted access to in-person clinical services and research activities, including data collection during the pandemic, speech data that can be collected remotely via phone or video calls became a promising avenue for research.² Overall, the findings illustrate a clear evolution from traditional, speech-based screening methods to more sophisticated, computationally powered diagnostic approaches. This transition not only demonstrates technological progress but also reflects a deeper understanding of dementia's complexity and the corresponding need for interdisciplinary methods to address it.

Our second key finding highlighted the convergence of clinical need and technological capability in topic areas. The keyword analysis revealed that most research in this line explored dementia-related disease (e.g., Alzheimer's disease) employing a specific methodology (e.g., speech analysis) in different subjects (e.g., human subjects of different ages). The predominant focus on dementia-related disease, particularly Alzheimer's disease, reflects the pressing clinical demand for accessible diagnostic tools in an aging global population. The World Health Organization's projection of a threefold increase in dementia cases by 2050 underscores the urgent need for developing accessible dementia diagnostic methods.²⁸ What is particularly significant is how this clinical need has been met with technological readiness—the availability of sophisticated tools such as Praat,²⁶ OpenSMILE,²⁷ ProsodyPro,²⁹ and deep learning frameworks³⁰ has enabled researchers to move beyond traditional manual speech analysis. This technological advancement has essentially removed a critical barrier to large-scale speech analysis, potentially accelerating the clinical application of research findings focusing on high-dimensional data analysis for classifying speech from dementia and nondementia groups.³¹

Finally, our abstract analysis identified a high correlation among certain terms (e.g., Alzheimer's disease and speech), illuminating the inherently interdisciplinary nature of this field. The correlation pattern suggests that research in this area is interdisciplinary with meaningful integration occurring across neuroscience, linguistics, and computer science. This integration is crucial because effective diagnosis through speech analysis requires understanding not only the neurological basis of dementia but also the linguistic manifestations of cognitive decline and the computational methods to detect these changes reliably.^32–34 Fraser and colleagues³⁵ demonstrated that advancements in machine learning and linguistic analysis would play a significant role in the assessment and classification of suspected Alzheimer's disease. Similarly, Fraser et al.³⁶ claimed that supplementing standard screening protocols with computerized analysis of language patterns could lead to more accurate assessments. Petti et al.³⁷ further emphasized how understanding the neurobiological basis of prosodic impairment helped inform which acoustic features were most diagnostically relevant. Interdisciplinary approaches that integrate neurological, linguistic, and computational methods, thus, may offer improved sensitivity and specificity in detecting dementia compared to methods focused on a single domain.

An unexpected but noteworthy finding, as a result of the abstract analysis, was the emergence of ALS in the co-occurrence analysis, despite it not being included in the initial search criteria. ALS is primarily a motor neuron disease, but it shares important overlaps with dementia-related conditions, particularly FTD. A growing body of clinical research has shown that cognitive and behavioral impairments are present in a substantial subset of ALS patients. For instance, Ceslis et al.³⁸ reported that approximately 14% of individuals with ALS have comorbid dementia, and around 31% demonstrate significant cognitive or language impairments. Up to 50% may experience mild-to-moderate cognitive or behavioral deficits. This intersection has led some researchers to include ALS in broader neurodegenerative comparisons or to use ALS cohorts when validating speech analysis tools developed for dementia detection. The co-occurrence of ALS with speech, diagnosis, and detection terms in our results likely reflects these patterns in the literature. In many cases, ALS appears not as a primary target but as a comparator condition or a contextual reference point in studies focused on dementia. This suggests that speech features relevant to cognitive assessment are sometimes explored across disease boundaries. Importantly, our study did not treat ALS as a target condition; rather, its presence emerged organically from the dataset, which was constructed solely using dementia-related search terms. The appearance of ALS, therefore, highlights how speech biomarkers are discussed across a spectrum of neurological conditions and emphasizes the interwoven nature of cognitive and motor impairments in the research discourse.

Despite increased research activity and technological advancement, our analysis also reveals several persistent gaps in the literature. First, there is a lack of consistency in how prosodic features are defined, labeled, and measured across studies, as evidenced by the wide variety of related terms, such as pitch and intonation, used throughout the literature we analyzed. Our co-occurrence analysis indicated that these terms often appear independently or in inconsistent clusters, suggesting a lack of standardized terminology. This inconsistency not only complicates the automated retrieval of relevant literature but also limits comparability and synthesis across studies and hinders the development of a unified diagnostic framework. Second, longitudinal investigations of speech and prosody across the course of dementia progression remain scarce, constraining insights into how prosodic changes evolve over time. Our analysis revealed a limited occurrence of terms related to temporal tracking or follow-up designs, indicating that most studies in this area rely on cross-sectional data. This scarcity constrains insights into how prosodic changes evolve over the course of a disease, which is critical for validating speech biomarkers as tools for early detection and disease monitoring. Third, research remains disproportionately focused on Alzheimer's disease, with less attention given to other dementia subtypes such as PPA or mixed dementia presentations. The heavy focus on Alzheimer's disease was revealed by our keyword and co-occurrence analysis and suggests that much of the current knowledge and methodological development is anchored in patterns specific to Alzheimer's disease, limiting the applicability of findings to broader dementia populations. In a nutshell, these gaps suggest that while the field is expanding, further conceptual and methodological refinement is needed to strengthen the clinical utility and generalizability of speech-based dementia biomarkers, supporting the findings of other researchers. ^8,39

The limitations of our study should be noted. First, our search strategy encompassed a wide range of dementia-related conditions, including those where motor impairments may substantially influence prosodic features (e.g., Parkinson's disease dementia, nonfluent/agrammatic PPA). These conditions may be associated with prosodic alterations that stem not only from cognitive decline but also from motor impairments, which can complicate interpretations of speech-related findings in the literature. While our study did not evaluate prosodic changes directly, this overlap may affect how prosody is conceptualized and measured across the field. Our object was to capture the breadth of research activity relating to speech and dementia, rather than isolate specific etiologies. Additionally, this study does not adhere to the full methodological rigor of a traditional systematic review. Although titles and abstracts were manually screened by two authors using predefined inclusion criteria, the process may be subject to selection bias. Furthermore, the text mining approach employed was intended to identify large-scale thematic and temporal patterns rather than to serve as a substitute for established bias control or critical appraisal procedures. Nonetheless, the findings of this study highlight an important challenge in the literature and suggest that future research trend analyses or systematic reviews may benefit from stratifying studies by the dementia subtype of underlying motoric versus cognitive mechanisms when feasible. Second, we also acknowledge that the use of broad acoustic terms in our search strategy may have captured studies focusing on nonprosodic features such as jitter, shimmer, and others. While efforts were made during screening to ensure studies have relevance to prosodic analysis, some overlap with general speech acoustic research may remain. This reflects the evolving and sometimes overlapping definitions of prosody and acoustic analysis in the literature. Future reviews with a narrower scope may consider applying more targeted inclusion criteria to isolate research specifically addressing prosodic phenomena. Second, the literature we analyzed may not consistently distinguish between prosodic changes driven by cognitive decline and those influenced by neuropsychiatric symptoms, such as apathy, depression, and anxiety. These symptoms are prevalent in individuals with Alzheimer's disease and related dementias, with estimates suggesting that over half of patients experience at least one neuropsychiatric symptom.⁴⁰ Moreover, such symptoms often emerge before overt cognitive decline and can independently impact prosodic and affective speech features.⁴¹ As our study focused on bibliometric trends rather than primary clinical data, some publications may implicitly attribute prosodic changes to neurodegeneration without accounting for these confounding factors. Future research on speech biomarkers should explicitly consider the role of neuropsychiatric symptoms to enhance diagnostic specificity. Third, our text mining approach was limited to peer-reviewed, English-language publications, potentially excluding relevant studies published in other languages or available in grey literature. This constraint may have influenced the comprehensiveness of topic representation across global research efforts. Expanding future reviews to include multilingual and unpublished sources could provide a more thorough picture of global research trends in this field.

Conclusion

Research on speech prosody as a potential biomarker for dementia is advancing rapidly, with a marked increase in publications since 2020 reflecting heightened interest and interdisciplinary momentum. Our findings suggest that the field is evolving toward broader integration of linguistic, neurological, and computational methods. However, this growth also underlines key challenges, including inconsistent terminology, limited longitudinal research, and an overemphasis on Alzheimer's disease relative to other dementia types. Addressing these gaps through methodological standardization, broader representation of dementia subtypes, and refined research focus will be critical for moving the field toward greater clinical relevance and utility.

Footnotes

ORCID iD

Chorong Oh

Min Sook Park

Contributorship

Conceptualization, C.O. and M. S. P.; methodology, C.O. and M. S. P.; software, M. S. P.; formal analysis, M. S. P.; investigation, C.O. and M. S. P.; resources, C.O. and M. S. P.; data curation, C.O. and M. S. P.; writing—original draft preparation, C.O.; writing—review and editing, C.O. and M. S. P.; visualization, M. S. P.; supervision, C.O. and M. S. P.; project administration, C.O. All authors have read and agreed to the published version of the manuscript.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability statement

The data supporting this study's findings are available from the corresponding author upon reasonable request.

References

Alzheimer’s Society. Types of dementia. Available at: https://www.alzheimers.org.uk/about-dementia/types-dementia (accessed 31 January 2025 ) .

König

Satt

Sorin

, et al. Use of speech analyses within a mobile application for the assessment of cognitive impairment in elderly people. Curr Alzheimer Res 2018; 15: 684–692.

Morris

Wang

, et al. Analysis of emotional prosody as a diagnostic tool for cognitive impairment. Front Psychol 2023; 14: 1–12.

Taler

Phillips

. Language performance in Alzheimer’s disease and mild cognitive impairment: a comparative review. J Clin Exp Neuropsychol 2008; 30: 501–556.

Roark

Mitchell

Hosom

J-P

, et al. Spoken language derived measures for detecting mild cognitive impairment. IEEE Trans Audio Speech Lang Process 2011; 19: 2081–2090.

Palmatier

Houston

Hulland

. Review articles: Purpose, process, and structure. J Acad Mark Sci 2018; 46: –5.

Boell

Cecez-Kecmanovic

. On being ‘systematic’ in literature reviews in IS. J Inf Technol 2015; 30: 161–173.

Morris

Wang

, et al. A systematic review of expressive and receptive prosody in people with dementia. J Speech Lang Hear Res 2021; 64: 3803–3825.

Ioannidis

JPA

. Why most clinical research is not useful. PLoS Med 2016; 13: e1002049.

10.

National Institute on Aging. What is dementia? Symptoms, types, and diagnosis. Available at: https://www.nia.nih.gov/health/alzheimers-and-dementia/what-dementia-symptoms-types-and-diagnosis (accessed 31 January 2025 ) .

11.

Mayo Clinic. Dementia. Available at: https://www.mayoclinic.org/diseases-conditions/dementia/symptoms-causes/syc-20352013 (accessed 31 January 2025 ) .

12.

Shattuck-Hufnagel

Turk

. A prosody tutorial for investigators of auditory sentence processing. J Psycholinguist Res 1996; 25: 193–247.

13.

Crowther

Lim

Crowther

. Systematic review and meta-analysis methodology. Blood 2010; 116: 3140–3146.

14.

Eden

Levit

Morton

(eds) Finding what works in health care: standards for systematic reviews. Washington (DC): National Academies Press, 2011.

15.

Huang

Yang

, et al. Detecting research topic trends by author-defined keyword frequency. Inf Process Manag 2021; 58: 102594.

16.

Tripathi

Kumar

Sonker

, et al. Occurrence of author keywords and keywords plus in social sciences and humanities research: a preliminary study. COLLNET J Scientomet Inf Manag 2018; 12: 215–232.

17.

Zhang

Zheng

, et al. Comparing keywords plus of WOS and author keywords: a case study of patient adherence research. J Assoc Inf Sci Technol 2016; 67: 967–972.

18.

Abdallah

Khalil

. Using a hybrid methodology for literature review: a case study in depression research. Inf Discov Deliv 2023; 52: 305–323.

19.

Jonnalagadda

Goyal

Huffman

. Automating data extraction in systematic reviews: a systematic review. Syst Rev 2015; 4: 4.

20.

Nie

Sun

. Using text mining techniques to identify research trends: a case study of design research. Appl Sci 2017; 7: 01.

21.

Feng

Chiam

SK.

Text-mining techniques and tools for systematic literature reviews: A systematic literature review. In: Proc. 24th Asia-Pacific Softw. Eng. Conf. (APSEC). IEEE. 2017. https://doi.org/10.1109/APSEC.2017.10.

22.

Higuchi

Koichi

. A two-step approach to quantitative content analysis: KH Coder tutorial using Anne of Green Gables (Part I). JRitsumeikan Social Sciences Review 2016; 52: 77–91.

23.

Higuchi

Koichi

. A Two-Step Approach to Quantitative Content Analysis: KH Coder Tutorial using Anne of Green Gables (Part II). Ritsumeikan Social Sciences Review. 2017; 53: 137–147.

24.

Feldman

Dagan

Knowledge discovery in textual databases. In: Proc. 1st Int Conf Knowl Discov Data Mining. AAAI Press, 1995, pp.112–117.

25.

Segev

. Semantic network analysis in social sciences. London: Routledge, 2021.

26.

Jaccard

. Étude comparative de la distribution florale dans une portion des alpes et des Jura. Bull Soc Vaudoise Sci Nat 1901; 37: 547–579.

27.

López-de-Ipiña

Alonso

Sole-Casals

, et al. On automatic diagnosis of Alzheimer's disease based on spontaneous speech analysis and emotional temperature. Cogn Comput 2013; 7: 44–55.

28.

World Health Organization. Dementia. Available at: https://www.who.int/news-room/fact-sheets/detail/dementia (accessed 31 January 2025 ) .

29.

Prosodypro–A tool for large-scale systematic prosody analysis. In: Tools and resources for the analysis of speech prosody. France: Laboratoire Parole et Langage, 2013, pp.7–10.

30.

Luz

Haider

de la Fuente

, et al. Alzheimer's dementia recognition through spontaneous speech: The ADReSS challenge. In: Proceedings of Interspeech 2020; 2020: 2172–2176. doi:https://doi.org/10.48550/arXiv.2004.06833

31.

Gosztolya

Bőhm

Vincze

, et al. Detecting mild cognitive impairment from spontaneous speech by combining automatically extracted prosodic, semantic, and syntactic features. Comput Speech Lang 2019; 53: 181–197.

32.

Chatzianastasis

Ilias

Askounis

, et al. Neural architecture search with multimodal fusion methods for diagnosing dementia. In: ICASSP 2023 IEEE Int Conf Acoustics, Speech and Signal Processing (ICASSP), 2023. doi:https://doi.org/10.1109/ICASSP49357.2023.10096579.

33.

Ding

Chetty

Noori Hoshyar

, et al. Speech-based detection of Alzheimer’s disease: a survey of AI techniques, datasets, and challenges. Artif Intell Rev 2024; 57: 1–43.

34.

Ilias

Askounis

. Multimodal deep learning models for detecting dementia from speech and transcripts. Front Aging Neurosci 2022; 14: 830943.

35.

Fraser

Lundholm Fors

Eckerström

, et al. Improving the sensitivity and specificity of MCI screening with linguistic information. In: Proc LREC 2018 Workshop Resources and Processing of Linguistic, Para-linguistic, and Extra-linguistic Data from People with Various Forms of Cognitive/Psychiatric Impairments (RaPID-2), 2015.

36.

Fraser

Meltzer

Rudzicz

, et al. Linguistic features identify Alzheimer’s disease in narrative speech. J Alzheimer’s Dis 2016; 49: 407–422.

37.

Petti

Baker

Korhonen

. A systematic literature review of automatic Alzheimer’s disease detection from speech and language. J Am Med Inform Assoc 2020; 27: 1784–1797.

38.

Ceslis

Argall

Henderson

, et al. The spectrum of language impairments in amyotrophic lateral sclerosis. Cortex 2020; 132: 349–360.

39.

Forbes-McKay

Shanks

Venneri

. Profiling spontaneous speech decline in Alzheimer’s disease: a longitudinal study. Acta Neuropsychol 2013; 25: 1–13.

40.

D’Onofrio

Sancarlo

Panza

, et al. Neuropsychiatric symptoms and functional status in Alzheimer’s disease and vascular dementia patients. Curr Alzheimer Res 2012; 9: 759–771.

41.

Ismail

Smith

Geda

, et al. Neuropsychiatric symptoms as early manifestations of emergent dementia: provisional diagnostic criteria for mild behavioral impairment. Alzheimers Dement 2016; 12: 195–202.

Unveiling 30 years of research on speech biomarker of dementia using text mining

Abstract

Introduction

Methods

Results

Discussion

Keywords

Introduction

Methods

Search strategy

Identification of studies

Data analyses

Results

Search outcomes

Temporal research trends

Key research topic areas and methodological approaches

Discussion

Conclusion

Footnotes

ORCID iD

Contributorship

Funding

Declaration of conflicting interests

Data availability statement

References