Alyssa N. Kaser1
Laura H. Lacritz1,2Holly R. Winiarski1
Peru Gabirondo3
Jeff Schaffert1Alberto J. Coca3,4Javier Jiménez-Raboso3Tomas Rojo3
Carla Zaldua3Iker Honorato3Dario Gallego3Emmanuel Rosario Nieves1,5
Leslie D. Rosenstein1,5
C. Munro Cullum1,2,6*- 1Department of Psychiatry, The University of Texas Southwestern Medical Center, Dallas, TX, United States
- 2Department of Neurology, The University of Texas Southwestern Medical Center, Dallas, TX, United States
- 3AcceXible Impacto, Sociedad Limitada, Bilbao, Spain
- 4Cambridge Mathematics of Information in Healthcare Hub, University of Cambridge, Cambridge, United Kingdom
- 5Parkland Health and Hospital System Behavioral Health Clinic, Dallas, TX, United States
- 6Department of Neurological Surgery, The University of Texas Southwestern Medical Center, Dallas, TX, United States
Objective: Early detection of cognitive impairment in the elderly is crucial for diagnosis and appropriate care. Brief, cost-effective cognitive screening instruments are needed to help identify individuals who require further evaluation. This study presents preliminary data on a new screening technology using automated voice recording analysis software in a Spanish population.
Method: Data were collected from 174 Spanish-speaking individuals clinically diagnosed as cognitively normal (CN, n = 87) or impaired (mild cognitive impairment [MCI], n = 63; all-cause dementia, n = 24). Participants were recorded performing four common language tasks (Animal fluency, alternating fluency [sports and fruits], phonemic “F” fluency, and Cookie Theft Description). Recordings were processed via text-transcription and digital-signal processing techniques to capture neuropsychological variables and audio characteristics. A training sample of 122 subjects with similar demographics across groups was used to develop an algorithm to detect cognitive impairment. Speech and task features were used to develop five independent machine learning (ML) models to compute scores between 0 and 1, and a final algorithm was constructed using repeated cross-validation. A socio-demographically balanced subset of 52 participants was used to test the algorithm. Analysis of covariance (ANCOVA), covarying for demographic characteristics, was used to predict logistically-transformed algorithm scores.
Results: Mean logit algorithm scores were significantly different across groups in the testing sample (p < 0.01). Comparisons of CN with impaired (MCI + dementia) and MCI groups using the final algorithm resulted in an AUC of 0.93/0.90, with overall accuracy of 88.4%/87.5%, sensitivity of 87.5/83.3, and specificity of 89.2/89.2, respectively.
Conclusion: Findings provide initial support for the utility of this automated speech analysis algorithm as a screening tool for cognitive impairment in Spanish speakers. Additional study is needed to validate this technology in larger and more diverse clinical populations.
Introduction
Alzheimer’s disease (AD) is the most common form of dementia and constitutes a significant public health problem worldwide due to the increasing need of social and economic services for patients and their support systems (1). A particular clinical challenge in AD and related dementias is identifying the early stages of cognitive decline due to its insidious onset and progression. AD is characterized by neuropathological features that can occur 15 to 20 years prior to noticeable changes in cognition or daily function (2, 3). Mild cognitive impairment (MCI) often represents a transitional stage between normal aging and dementia, and these individuals are at a greater risk of developing incident AD (4–8). Detecting early stages of cognitive decline is paramount for identifying those at greatest risk of progression to dementia and who may benefit from further evaluation and treatment (9). Improving the detection of early cognitive decline may lead to earlier initiation of disease modifying treatments to preserve independent functioning, which is crucial to a patient’s quality of life.
There currently exist numerous tools to help detect early clinical or pathological signs of MCI/dementia, including thorough clinical workups, neuropsychological evaluations, neuroimaging, and blood and cerebrospinal fluid biomarkers. While imaging and biomarker-based techniques can detect some neuropathological features associated with AD, some are invasive and not widely accessible for the purpose of early detection, and others are still in development. Formal neuropsychological evaluation is often an intrinsic part of the diagnostic process for detecting subtle changes in cognitive function in many centers. However, access to neuropsychological services may be limited due to various geographic, economic, and/or psychosocial factors, including limited access to healthcare resources, and therefore cannot be utilized effectively on a large scale. For these reasons, cognitive screening instruments are used as brief and cost-effective methods to identify individuals who require further evaluation and can be implemented across both primary care and specialized clinical settings. Primary care providers often serve as the first line of medical care for older adults concerned about their cognitive performance, yet physicians in primary care settings report feeling ill-equipped to recognize or document signs of cognitive decline (10, 11), contributing to gross under-detection of MCI in primary care settings (12). Nonetheless, current screening methods are not sufficiently sensitive or reliable in identifying clinical signs in the very early stages of dementia (13), and more precise screening techniques that can capture subtle changes in cognition are needed to identify individuals who may benefit from further evaluation.
Novel technologies and automated software systems to assess cognitive functioning in older individuals are emerging as new methods for early detection. In recent years, artificial intelligence (AI) and machine learning (ML) methods have surfaced as promising tools to aid in the early detection and diagnosis of AD and related disorders (14). More specifically, several studies have proposed the use of linguistic biomarkers for clinical classification and screening purposes (15–17). Identification and classification of language abnormalities play an important role in the diagnosis of AD, as subtle changes in various aspects of language have been identified in the early stages of the disease, including changes in number of between-utterance pauses (18), verbal fluency (19) and confrontation naming (20). Language impairment is typically defined by difficulties with word finding, comprehension, naming, and/or spontaneous speech. Simple tasks such as naming or verbal fluency can capture changes in language, and these markers of language impairment have been associated with early cognitive impairment in various neurodegenerative disorders and in some cases may precede other diagnostic clinical features (21, 22). However, these measures focus on the content of language rather than distinctive linguistic processes that may also be impacted in neurodegenerative disorders, such as speech rate, frequency and duration of pauses, and discourse efficiency (23). Furthermore, language tasks involve other cognitive processes that are relevant to the onset of dementia, including verbal working memory, attention, and processing speed (24). Therefore, sophisticated speech analyses might provide clinically relevant information, pertaining to both linguistic functions and different aspects of cognition (25). Moreover, the relative ease with which linguistic information can be accessed in a clinical context further favors the use of speech biomarkers as an accessible early detection tool.
In fact, machine learning (ML) techniques have been developed utilizing various speech and linguistic biomarkers to identify individuals with MCI (26), early AD (27, 28), dementia (29), Parkinson’s disease (30), and frontotemporal disorders (31). For instance, Hajjar and colleagues (27) found both acoustic and lexical-semantic biomarkers to be sensitive to cognitive impairment and disease progression in the early stages of AD. Several additional studies have demonstrated efficacy of computer-assisted linguistic processing algorithms in detecting speech differences between normal and impaired individuals in the English language (32, 33). Similar efficacy has been illustrated in other languages, including Chinese (34), Hungarian (26) and Swedish (35). While dementia-related linguistic changes in Spanish speakers continue to be an emerging area of research (36–38, 39), they remain understudied. Few investigations have explored automated speech analysis for detection of cognitive decline in a Spanish speaking sample (40, 41), and even so, published findings have limited interpretability for clinical use. Furthermore, several of these studies have been rated for high risk of bias due to sample selection (42), as measured by the QUADAS-2 checklist for quality assessment of diagnostic accuracy studies (43), potentially limiting its clinical applicability.
According to the Alzheimer’s Association, Hispanic individuals are disproportionately more likely to develop Alzheimer’s disease and related dementias (44) than their non-Hispanic White counterparts. Furthermore, this population is estimated to have the largest projected increase in ADRD over the next several decades (45). However, older adults in the Hispanic community frequently face challenges in receiving timely and accurate diagnoses. This is often attributed to the limited availability of bilingual providers and culturally validated neuropsychological tests, compounded by systemic barriers that hinder access to high-quality healthcare and specialized services (46). Survey findings revealed that Hispanic Americans rely on primary care or community health providers to screen for and address cognitive concerns, though standardized cognitive screening remains variable in such settings. Furthermore, in a scoping review of available cognitive screening tools for Spanish speaking populations, Burke and colleagues (47) suggested that some of the most frequently used cognitive screening tools may be less valid in this heterogenous population. As such, developing inclusive and culturally sensitive screening tools to detect cognitive impairment in this and other ethnic groups is crucial.
In the present study, we aimed to contribute to this area of exploration by presenting preliminary data on a novel cognitive screening technology that uses automated linguistic analysis software to quickly screen for MCI and dementia in a Spanish population. Furthermore, this study addresses a gap in the literature by introducing a brief, cost-effective, and readily available screening tool via web-based platform that is fully automated, highlighting its ease and accessibility in primary care and community health settings. The goals of the current manuscript were to (1) Validate a machine learning algorithm in a training sample, and (2) Test this algorithm in an aging Spanish-speaking testing sample to validate model performance when comparing cognitively normal (CN) with impaired (MCI + dementia) and MCI groups.
Method
Sample
Data were collected from 195 Spanish-speaking individuals recruited as part of a Multicenter Clinical Trial in Madrid, Getxo, and Quiron Salud, Spain. Participants were assessed by a neuropsychologist and neurologist and completed a series of cognitive tests as part of the initial evaluation: The Mini Mental State Examination (MMSE) (48), the 7-min screen test (49), Clock Drawing Test (50), Trail Making Test Parts A and B (51), the Blessed Dementia Scale (52), and four verbal tasks described below. Clinical diagnosis of normal cognition, MCI, or dementia was made by participating neurologists and neuropsychologists based on a full clinical work-up that included neurological and neuropsychological evaluation (without consideration of the speech processing variables). For a diagnosis of MCI, Petersen criteria (53) were utilized, and all subtypes of MCI were included (i.e., amnestic; multiple-domain; single-domain, non-amnestic). All-cause dementia was diagnosed in accordance with the DSM-IV criteria (54), requiring a notable decline in complex instrumental activities of daily living. Participants were considered eligible for the study if they (1) were proficient in the Spanish language, and (2) had been clinically diagnosed with either cognitive impairment (e.g., MCI or dementia) (7) or considered to have normal cognition. Exclusion criteria included prior diagnosis of a significant psychiatric disorder, cognitive impairment not due to a neurodegenerative process, and significant visual, hearing, or expressive language impairment that would impact the ability to complete cognitive tasks. Data from 21 participants were discarded due to incomplete cognitive data (n = 8), missing demographic information (n = 10), or poor text transcription quality on the audio recording used for algorithm modeling (n = 3). The sample was divided into a training set (N = 122), and a validating/testing set (N = 52; see Table 1 for sample characteristics).
Table 1. Demographic characteristics of the sample.
Measures
Cognitive assessments
The current study utilized four language tasks for capturing neuropsychological performance and audio recording variables for each participant, including (1) Picture Description Task (PDT), using the “Cookie Theft” picture from the Boston Diagnostic Aphasia Examination which assesses spoken language (55); (2) Phonemic Verbal Fluency (PVF), measured by the letter (F) word generation task; (3) Alternating Verbal Fluency (SVF-Alt), measured by alternating fruits and sports word generation task; and (4) Semantic Verbal Fluency (SVF), measured by animal naming fluency (see Figure 1). Tasks were administered and audio files were recorded and stored through the AcceXible platform, a proprietary instrument developed for disease detection through speech analysis. AcceXible is a web-based platform that can be accessed via web browser or mobile application designed for Apple® devices, and utilizes cloud-based data processing and storage features (56; see Supplementary Figure S1). Once the initial task was initiated by the clinician or administrator, participants followed instructions presented on the computer interface for each of the four tasks. Total administration time was approximately 5 min.
Figure 1. Five classifiers combined to create machine learning algorithm.
Acoustic and linguistic feature extraction
Recordings were processed via speech-to-text transcription (Google Cloud’s Speech-to-Text) and digital-signal processing techniques to capture neuropsychological variables and audio characteristics, respectively. SVF, SVF-Alt and PVF tasks were restricted to 1 min each to ensure that all audio files had a uniform duration of 1 min (mean length of ~61 s). Due to the nature of the task, the Picture Description Task (PDT) was not limited based on time, and audio recording concluded based on provider’s clinical judgment. Audio recordings were segmented into 25 ms windows, a practice commonly seen in audio segmentation literature (Supplementary Table S1) (57). Audio segments were then preprocessed by normalizing the amplitude, to minimize bias toward signals with higher or lower energy levels, and using Butterworth low-pass filter, to attenuate high frequencies and eliminate impact of environmental noise (58, 59). Variables were computed and used for voice model training using the Python library “librosa” (60). Extracted variables from verbal tasks included consecutive repetitions, mean duration and standard deviation of silences, number of repeated words, and other task-specific variables. An exhaustive list of extracted speech features and verbal task variables used for model training can be found in Tables 2, 3.
Table 2. Verbal task variables extracted for model training.
Table 3. Speech Feature Variables Extracted for Model Training.
Algorithm development and model training
Clinical diagnosis served as the target variable for training the ML model. In the current study, this variable contained three diagnostic categories: CN, MCI, and dementia. The model considered the objective variable as dichotomous, with a score of zero indicating cognitively normal and a score of one indicating impairment (MCI + dementia or MCI, depending on the impairment group of interest). The model’s objective was to identify the earliest stages of cognitive decline, as well as overt dementia. As previously mentioned, the initial dataset was separated into a training and testing set using a 70%/30% split to ensure a socio-demographically balanced training sample. The training procedure was applied to the training set and the testing set was used to test the final model. To train a model that had the same predictive power when including sociodemographic information, a demographically similar subsample (age, sex, education, and clinical center) across groups of diagnoses was used.
First, the linguistic and acoustic features were extracted and preprocessed using text transcription techniques and digital-signal preprocessing, respectively. In the case of acoustic features (see Table 3), a feature selection step was added in each iteration of the Cross Validation due to the high number of features. The feature selection was based on the F-statistic value obtained in the analysis of variance. The number of selected features was added as a hyperparameter in the Grid Search CV, so that the best “k” features were selected based on F-statistic results, given a grid (5, 10, 15, 20). In other cases (e.g., SVF, SVF-Alt, PDT, and PVF), all computed features described in Table 2 were used to train the ML models. Features were standardized before building the models. The detailed descriptions of these variables are presented in Tables 2, 3. For each variable described in Table 2, including neuropsychological variables, a Python-based function was defined to perform its extraction and calculation, allowing for automatic computation when the transcription was carried out (61–63). All predictive models used in this study were classic ML models registered in scikit-learn software (64). These models include Logistic Regression, Supported Vector Machines, K-Nearest Neighbors, Random Forest and Gradient Boosting. The strategy used to build models was Grid Search 10-fold Cross Validation (CV), where model selection was performed by maximizing AUC score (Supplementary Data).
Subsequently, five independent ML models were trained: one for each task, using the extracted linguistic variables and another one with the digital-signal processing features, as demonstrated in Figure 2. When building the model based on acoustic features, a Support Vector Machine (SVM) with a linear kernel and regularization strength of 10 was used. A feature selection step was incorporated to address the high number of acoustic features, with the optimal number of features determined to be 15 by the SVM (see Supplementary Table S2). Each one of the five independent models followed the same training procedure. A 10-fold CV was performed 10 times to increase robustness and avoid overfitting. The algorithm development procedure began by first defining a list of candidate ML classical models and identifying their respective hyperparameter grids. For each model candidate, the sample was randomly stratified into a 90/10 split, a commonly used train/test ratio when implementing a 10-fold CV procedure (65). Next, model parameters were calculated on the training set and scores were calculated for the new testing set, ranging from zero to one. Sensitivity, specificity, accuracy, area under the curve (AUC), and mean squared error (MSE) were calculated for each possible cutoff point from zero to one. These steps were repeated 10 times and the mean of each calculated metric was reported, at which point this step was completed 10 more times to report the mean and variance of each obtained result. The model and cutoff point that demonstrated superior performance based on all calculated metrics was then manually selected. Finally, the parameters of the selected model were calculated for the training set sample. The threshold was chosen during the training phase. For each candidate threshold (ranging from 0 to 1 in increments of 0.01), we computed the mean accuracy, mean sensitivity, and mean specificity using a 10-fold CV approach. The optimal threshold point for detecting impaired (MCI + dementia) and MCI groups was determined by (1) maximizing the sum of mean sensitivity and mean specificity by way of Youden’s J statistic (66) [calculated as (sensitivity + specificity – 1)] and (2) ensuring that the mean sensitivity exceeded the mean specificity at that point. If multiple threshold points met these conditions, the smallest threshold value was selected to prioritize sensitivity over specificity.
Figure 2. Model training and testing scheme.
Once this procedure was performed for the various models (see Figures 1, 2), the predicted scores were computed for each model (i.e., four tasks and one speech) to create five individual scores for each participant. Finally, an ensemble ML model was trained in which the input included the five individual scores, and the output was one unique score. Before training the ensemble model, logit transformation was applied to each obtained score to create a continuous value in the range of negative infinity to infinity to obtain a final model.
Statistical analyses
Demographic characteristics for the training and testing samples are presented in Table 1. Comparisons between cognitively intact and impaired groups were calculated using two sample t-test in case of continuous variables (i.e., age, education, MMSE) and chi-square test for proportions in case of categorical variables (i.e., sex, clinical center). General linear modeling was used to test whether age, sex, and education significantly predicted logistically transformed algorithm scores in the training and testing sets. Receiver operating characteristic (ROC) curves were constructed for the testing set when discriminating CN and impaired (MCI + dementia) or MCI groups, and areas under the curves (AUC) were used as measures of overall classification accuracy. Accuracy, sensitivity, specificity, and F-scores were calculated for various threshold values for impaired and MCI groups. F-scores represent a common machine learning metric used to measure model accuracy by way of balancing precision and recall, with precision representing the proportion of correct “positive” predictions made by the model, or positive predictive value, and recall representing the proportion of actual positive samples correctly identified by the model, or sensitivity (67). An F-score value ranges from 0 to 1, with a 1 indicating perfect precision and recall. ANCOVA was used to examine differences in mean logit and algorithm scores by CN, MCI, and all-cause dementia groups, controlling for age, sex, and education. Before performing this test, logit transformation was applied to better meet the assumptions of the test, since the outcome variable might be the prediction of a non-linear ML model, and residuals might show a non-constant spread. Finally, we computed accuracy, sensitivity, and specificity across different MMSE score thresholds for both impaired and MCI groups to compare the classification accuracy with that of the presented model. Statistical analyses were conducted using the statsmodels and pingouin packages in Python.
Results
Descriptive statistics
Demographic and clinical data of the 174 participants are described in Table 1. The cohort had a mean age of 74.03 years (SD = 8.55), completed an average of 12.45 years of education (SD = 4.72), and included slightly more females than males (56%). Demographic characteristics were similar across groups in the training set, but age and sex were significantly different between clinical groups in the testing sample. MMSE scores were different across groups in both samples (Supplementary Figure S2). A one-way ANCOVA, controlling for age, sex, and years of education, demonstrated a significant difference in mean logit and algorithm scores for every independent model’s prediction across CN, MCI, and all-cause dementia groups, except for phonemic verbal fluency (see Table 4).
Table 4. Mean logit and algorithm scores by clinical diagnosis in the testing sample.
Model performance
The final algorithm in the testing sample with the presented tool resulted in an AUC of 0.93 with an overall accuracy of 88.4%, sensitivity of 87.5%, specificity of 89.2%, and F-score of 0.87 in discriminating CN and impaired groups. The final algorithm in the testing sample obtained an AUC of 0.90 with an overall accuracy of 87.5%, sensitivity of 83.3%, specificity of 89.2%, and F-score of 0.80 in discriminating CN and MCI groups. Using Youden’s Index, the optimal threshold value was determined to be 0.45 for impaired (J = 0.767) and MCI (J = 0.725) groups. Figures 3, 4 present accuracy, F-score, sensitivity and specificity for each threshold value when discriminating CN versus impaired and MCI groups in the whole testing set. Since the selected optimal ensemble ML model is RandomForest, the distribution of scores was non-normal and scores fell close to 0 or 1. Therefore, metrics remained constant between the thresholds of 0.23 and 0.70. The mean squared error obtained in the testing set was 0.10 when distinguishing impaired from non-impaired and 0.114 when distinguishing MCI from non-impaired.
Figure 3. Accuracy, F-score, sensitivity and specificity for each threshold value in discriminating cognitively normal and impaired groups in the testing set.
Figure 4. Accuracy, F-score, sensitivity and specificity for each threshold value in discriminating cognitively normal versus mild cognitive impairment groups in the testing set.
Table 5 presents the performance of each machine learning model in the testing sample. Accuracy, sensitivity, and specificity were identical across SVF and ensemble models. Nonetheless, the final ensemble model demonstrated the best overall performance (AUC = 0.93) compared to each individual ML model, with the speech-only model performing the worst (AUC = 0.64). For comparison, when differentiating impaired from non-impaired groups, the MMSE demonstrated an accuracy range of 71.1 to 80.7%, sensitivity range of 41.6 to 66.6%, and specificity range of 75 to 100% across cut-offs of 24 to 28, with an optimal cut-off of 27 yielding an accuracy of 76.9%, sensitivity of 66.6%, and specificity of 85.7%. When differentiating MCI from non-impaired, the MMSE demonstrated an accuracy range of 62.5 to 77.5%, sensitivity range of 8.3 to 33.3%, and specificity range of 75 to 100% across cut-offs of 24 to 28, with an optimal cut-off of 27 yielding an accuracy of 70%, sensitivity of 33.3%, and specificity of 85.7%.
Table 5. Machine learning models’ performance in the testing sample for cognitively normal versus impaired and MCI groups.
Discussion
Novel automated technologies are emerging to allow methods to screen for subtle cognitive changes in older adults, as early detection is critical for diagnosis and initiation of appropriate care. This is especially important for historically marginalized racial and ethnic groups, who often experience delayed diagnosis of cognitive disorders and face barriers in accessing specialists for evaluation and treatment. Furthermore, these evolving technologies hold promise as brief, accessible, and scalable means of capturing emerging cognitive deficits. The current study explored the utility of an automated speech analysis algorithm to quickly and effectively screen for cognitive impairment in a Spanish speaking population.
Model performance
This speech analysis algorithm was able to accurately differentiate CN from impaired (MCI + dementia) and MCI groups with an overall accuracy of 88.4 and 87.5% in the testing set, respectively (Table 5). Examination of the associated AUC values of CN versus impaired and MCI groups in Figure 5 reveals close similarities between classification curves, both falling in the outstanding range (0.93 for impaired and 0.90 for MCI). Furthermore, F-scores for the final algorithm in both impaired and MCI groups were considered good, indicating that both precision and recall of the ML model were high. Therefore, this automated speech analysis algorithm holds promise for distinguishing both early and more advanced cognitive decline from those who are cognitively normal in the studied sample. Prior efforts toward developing speech analysis tools for early detection of cognitive impairment have demonstrated similar differentiating abilities, with initial results exhibiting high accuracy (80–95%) in the detection of subtle changes in speech of those diagnosed with MCI and AD (42). Previous ML studies have demonstrated promising diagnostic ability when using similar verbal tasks in various languages, including spontaneous speech (68), picture description task (69), and fluency tasks (70), in combination with acoustic and linguistic features of speech. The current approach adds to the literature by utilizing numerous verbal tasks in one ML model for a Spanish-speaking population, while simultaneously maintaining a brief administration time, though future research may clarify which task and/or tasks contribute most to the success of the model. For instance, given that the semantic verbal fluency model with associated verbal characteristics performed similarly to the final ensemble model in the current study, abbreviated versions of this task with similar screening abilities may be possible and should be further explored.
Figure 5. Receiver operating characteristic (ROC) curves and associated AUC values for cognitively normal (CN) versus mild cognitive impairment group, and CN versus impaired (MCI + dementia) group in the testing set.
Strengths, limitations, and future directions
The current study provides support for the utility of a quick, 5-minute assessment that analyzes responses to language measures to investigate aspects of speech that may not be easily detectable to examiners. Early screening and detection of cognitive impairment in at-risk populations may allow for the proper allocation of time and resources to individuals identified as requiring more extensive assessment (e.g., biomarker analysis, comprehensive neuropsychological evaluation, neuroimaging, etc.). Additionally, this speech analysis tool may outperform current standard-of-care screening measures such as the Mini-Mental State Examination. In this study, the technology outperformed the MMSE in screening for MCI and dementia among Spanish speakers. The tool demonstrated superior accuracy (87.5/88.4), sensitivity (83.3%/87.5%), and specificity (89.2%/89.2%) when compared to an MMSE cut-off of 24 or 25, two thresholds well-established in clinical practice (71). While qualitatively different, the current model’s performance seems comparable, if not superior, to the MMSE.
Blesa and colleagues developed age and education-adjusted cut-off scores for the MMSE in a Spanish speaking cohort, identifying the optimal cut-off score to be 24/25 which yielded a sensitivity of 87.32% and specificity of 89.19% in detecting dementia (72). Similar results have been demonstrated using the Montreal Cognitive Assessment (MoCA) in the detection of MCI among Spanish speakers, with a sensitivity of 80.48% and specificity of 81.19% (73). Beyond its superior sensitivity, this technology fills a gap in the existing literature as a concise, cost-effective screening tool that is available through a web-based platform. The tool is fully automated, partially addressing issues related to the frequent incongruence between the patient and examiner’s native language, and ultimately emphasizing its ease and accessibility in primary care and community health settings. Furthermore, the impact of the current model is a promising avenue to increase cognitive screening, as it could be easily adapted to mobile phones, the feasibility and efficacy of which has been demonstrated in prior studies (74, 75). Thus, the current technology may allow for screening of more individuals that may be otherwise unable to access care.
Despite the promising findings for advancing the detection and monitoring of cognitive impairment, certain limitations must be noted. First, the current state of machine learning research emphasizes the need for standardization of training set size determination. Future studies would benefit from an increased sample size to improve robustness and minimize bias of the clinical model. Second, speech analysis software relies on high-quality audio recording and transcription to ensure accurate interpretation of speech. This study experienced varied audio-quality during data collection, requiring three cases with poor quality to be excluded from analyses. Nonetheless, speech-to-text transcription models should continue to be updated with technological advances to further improve efficacy in future machine learning and technology studies. Third, this study included only Spanish speaking individuals from selected sites in Spain, and thus, the generalizability of the algorithm in other populations will require additional validation. To extend the utility of the current technology, future research may focus on other Spanish-speaking populations from various geographic locations which could represent dialectical differences. To this point, preliminary data utilizing this technology with a U.S.-based sample of Spanish speakers demonstrated the impact subtle variations in Spanish expressive language may have on accuracy of transcriptions (76). However, modification of the transcription tool to include these language variations resulted in an overall transcription accuracy of 95%. Therefore, differences in cultural exposure and regional dialect should be considered in the development and use of automated speech software.
While certain aspects of the current model are inherently more naturalistic than simple quantitative assessments of speech production, it can be argued that elements of the task are not entirely naturalistic (e.g., listing animals). Furthermore, the speech samples are somewhat constrained by the assessment setting, and thus may not accurately reflect an examinee’s naturalistic speech qualities. Various forms of spontaneous speech have shown to be sensitive to cognitive decline, including analysis of informal conversations with others (77) and detailed recollections of the preceding day’s events (78) or patients’ daily routine (79). Therefore, further research would be helpful in determining how the presented model’s assessment of these variables compares between the test setting and a more natural speech interaction, and how that might influence clinical interpretation.
Though speech analysis technology would provide an affordable and accessible option for identification of cognitive impairment, it is critical that investigations integrate biomarkers, more detailed neuropsychological evaluations, and neuroimaging findings to demonstrate incremental validity and increase our confidence in the clinical models. Additionally, as promising as the current model appears, future research may focus on the utility of the tool in predicting those at risk of conversion from MCI to dementia. For example, current stand-alone single-administration screening tools such as the MMSE are not effective for this use and may depend upon serial assessments, ideally with more comprehensive evaluations (80). Similarly, the tool may be further validated for use in differential diagnosis of neurodegenerative and other neurological conditions. A critical review on the role of connected speech in neurodegenerative diseases revealed numerous studies highlighting unique speech features in clinical populations, including amnestic MCI, primary progressive aphasia, movement disorders, dementia due to Alzheimer’s disease, and amyotrophic lateral sclerosis (23). Such detailed assessments of speech samples may add further value to traditional language assessment instruments.
Conclusion
The current findings provide initial support for the utility of an automated speech analysis algorithm to quickly and efficiently detect cognitive impairment in an older Spanish-speaking population. Results suggest that the algorithm has a robust ability to detect early stages of cognitive decline and clear dementia. Further research is needed to validate this methodology in additional languages and clinical populations, as this may be a valuable cross-cultural screening method for MCI and dementia.
Data availability statement
The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.
Ethics statement
The studies involving humans were approved by the CEIC Hospital Clínico San Carlos. The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.
Author contributions
AK: Writing – original draft. LL: Writing – review & editing. HW: Writing – original draft. PG: Writing – original draft. JS: Writing – review & editing. AC: Writing - original draft. JJ-R: Writing – review & editing. TR: Writing – review & editing. CZ: Writing – review & editing. IH: Writing – review & editing. DG: Writing – review & editing. EN: Writing – review & editing. LR: Writing – review & editing. CC: Writing – review & editing.
Funding
The author(s) declare that no financial support was received for the research, authorship, and/or publication of this article.
Conflict of interest
PG, AC, JJ-R, TR, CZ, IH, and DG were employed by Accexible.
The remaining 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.
Publisher’s note
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.
Supplementary material
The Supplementary material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fneur.2024.1342907/full#supplementary-material
References
1. Alzheimer’s Association. Alzheimer’s disease facts and figures. Alzheimers Dement. (2023) 19:1598–695. doi: 10.1002/alz.13016
2. Bateman, RJ, Xiong, C, Benzinger, TLS, Fagan, AM, Goate, A, Fox, NC, et al. Clinical and biomarker changes in dominantly inherited Alzheimer’s disease. N Engl J Med. (2012) 367:795–804. doi: 10.1056/NEJMoa1202753
3. Dubois, B, Hampel, H, Feldman, HH, Scheltens, P, Aisen, P, Andrieu, S, et al. Preclinical Alzheimer’s disease: definition, natural history, and diagnostic criteria. Alzheimers Dement. (2016) 12:292–323. doi: 10.1016/j.jalz.2016.02.002
4. Dubois, B, and Albert, ML. Amnestic MCI or prodromal Alzheimer’s disease? Lancet Neurol. (2004) 3:246–8. doi: 10.1016/S1474-4422(04)00710-0
5. Gauthier, S, Reisberg, B, Zaudig, M, Petersen, RC, Ritchie, K, Broich, K, et al. Mild cognitive impairment. Lancet. (2006) 367:1262–70. doi: 10.1016/S0140-6736(06)68542-5
6. Jessen, F, Wolfsgruber, S, Wiese, B, Bickel, H, Mösch, E, Kaduszkiewicz, H, et al. AD dementia risk in late MCI, in early MCI, and in subjective memory impairment. Alzheimers Dement. (2014) 10:76–83. doi: 10.1016/j.jalz.2012.09.017
7. Petersen, RC, Lopez, O, Armstrong, MJ, Getchius, TSD, Ganguli, M, Gloss, D, et al. Practice guideline update summary: mild cognitive impairment: report of the guideline development, dissemination, and implementation Subcommittee of the American Academy of neurology. Neurology. (2018) 90:126–35. doi: 10.1212/WNL.0000000000004826
8. Ward, A, Tardiff, S, Dye, C, and Arrighi, HM. Rate of conversion from prodromal Alzheimer’s disease to Alzheimer’s dementia: a systematic review of the literature. Dement Geriatr Cogn Dis Extra. (2013) 3:320–32. doi: 10.1159/000354370
9. Albert, MS, DeKosky, ST, Dickson, D, Dubois, B, Feldman, HH, Fox, NC, et al. The diagnosis of mild cognitive impairment due to Alzheimer’s disease: recommendations from the National Institute on Aging-Alzheimer’s association workgroups on diagnostic guidelines for Alzheimer’s disease. Alzheimers Dement. (2011) 7:270–9. doi: 10.1016/j.jalz.2011.03.008
10. Bradford, A, Kunik, ME, Schulz, P, Williams, SP, and Singh, H. Missed and delayed diagnosis of dementia in primary care: prevalence and contributing factors. Alzheimer Dis Assoc Disord. (2009) 23:306–14. doi: 10.1097/WAD.0b013e3181a6bebc
11. Mitchell, AJ, Meader, N, and Pentzek, M. Clinical recognition of dementia and cognitive impairment in primary care: a meta-analysis of physician accuracy. Acta Psychiatr Scand. (2011) 124:165–83. doi: 10.1111/j.1600-0447.2011.01730.x
12. Liu, Y, Jun, H, Becker, A, Wallick, C, and Mattke, S. Detection rates of mild cognitive impairment in primary care for the United States Medicare population. J Prev Alz Dis. (2023) in press) 11:7–12. doi: 10.14283/jpad.2023.131
13. Petrazzuoli, F, Vestberg, S, Midlöv, P, Thulesius, H, Stomrud, E, and Palmqvist, S. Brief cognitive tests used in primary care cannot accurately differentiate mild cognitive impairment from subjective cognitive decline. J Alzheimers Dis. (2020) 75:1191–201. doi: 10.3233/JAD-191191
14. Li, R, Wang, X, Lawler, K, Garg, S, Bai, Q, and Alty, J. Applications of artificial intelligence to aid early detection of dementia: a scoping review on current capabilities and future directions. J Biomed Inform. (2022) 127:104030. doi: 10.1016/j.jbi.2022.104030
15. Eyigoz, E, Mathur, S, Santamaria, M, Cecchi, G, and Naylor, M. Linguistic markers predict onset of Alzheimer’s disease. EClinicalMedicine. (2020) 28:100583. doi: 10.1016/j.eclinm.2020.100583
16. Orimaye, SO, Wong, JS-M, Golden, KJ, Wong, CP, and Soyiri, IN. Predicting probable Alzheimer’s disease using linguistic deficits and biomarkers. BMC Bioinformatics. (2017) 18:34. doi: 10.1186/s12859-016-1456-0
17. Tavabi, N, Stück, D, Signorini, A, Karjadi, C, Al Hanai, T, Sandoval, M, et al. Cognitive digital biomarkers from automated transcription of spoken language. J Prev Alzheimers Dis. (2022) 9:791–800. doi: 10.14283/jpad.2022.66
18. Pistono, A, Jucla, M, Barbeau, EJ, Saint-Aubert, L, Lemesle, B, Calvet, B, et al. Pauses during autobiographical discourse reflect episodic memory processes in early Alzheimer’s disease. J Alzheimers Dis. (2016) 50:687–98. doi: 10.3233/JAD-150408
19. Rodriguez-Aranda, C, Waterloo, K, Johnson, SH, Eldevik, P, Sparr, S, Wikran, GC, et al. Neuroanatomical correlates of verbal fluency in early Alzheimer’s disaese and normal aging. Brain Lang. (2016) 155-156:24–35. doi: 10.1016/j.bandl.2016.03.001
20. Cerbone, B, Massman, PJ, Woods, SP, and York, MK. Benefit of phonemic cueing on confrontation naming in Alzheimer’s disease. Clin Neuropsychol. (2020) 34:368–83. doi: 10.1080/13854046.2019.1607904
21. Mueller, KD, Koscik, RL, Hermann, BP, Johnson, SC, and Turkstra, LS. Declines in connected language are associated with very early mild cognitive impairment: results from the Wisconsin registry for Alzheimer’s prevention. Front Aging Neurosci. (2018) 9:437. doi: 10.3389/fnagi.2017.00437
22. Verma, M, and Howard, RJ. Semantic memory and language dysfunction in early Alzheimer’s disease: a review. Int J Geriatr Psychiatry. (2012) 27:1209–17. doi: 10.1002/gps.3766
23. Boschi, V, Catricala, E, Consonni, M, Chesi, C, Moro, A, and Cappa, SF. Connected speech in neurodegenerative language disorders: a review. Front Psychol. (2017) 8:269. doi: 10.3389/fpsyg.2017.00269
24. Kempler, D, and Goral, M. Language and dementia: neuropsychological aspects. Annu Rev Appl Linguist. (2008) 28:73–90. doi: 10.1017/S0267190508080045
25. Szatloczki, G, Hoffmann, I, Vincze, V, Kalman, J, and Pakaski, M. Speaking in Alzheimer’s disease, is that an early sign? Importance of changes in language abilities in Alzheimer’s disease. Front Aging Neurosci. (2015) 7:195. doi: 10.3389/fnagi.2015.00195
26. Toth, L, Hoffmann, I, Gosztolya, G, Vincze, V, Szatloczki, G, Banreti, Z, et al. A speech recognition-based solution for the automatic detection of mild cognitive impairment from spontaneous speech. Curr Alzheimer Res. (2018) 15:130–8. doi: 10.2174/1567205014666171121114930
27. Hajjar, I, Okafor, M, Choi, JD, Moore, E, Abrol, A, Calhoun, VD, et al. Development of digital voice biomarkers and associations with cognition, cerebrospinal biomarkers, and neural representation in early Alzheimer’s disease. Alzheimers Dement (Amst). (2023) 15:e12393. doi: 10.1002/dad2.12393
28. Campbell, EL, Docio-Fernandez, L, Jiménez-Raboso, J, and Gacia-Mateo, C. Alzheimer's dementia detection from audio and language modalities in spontaneous speech. Proc IberSPEECH. (2021) 2021:270–4. doi: 10.21437/IberSPEECH.2021-57
29. Xue, C, Karjadi, C, Paschalidis, IC, Au, R, and Kolachalama, VB. Detection of dementia on voice recordings using deep learning: a Framingham heart study. Alzheimers Res Ther. (2021) 13:146. doi: 10.1186/s13195-021-00888-3
30. Favaro, A, Moro-Velázquez, L, Butala, A, Motley, C, Cao, T, Stevens, RD, et al. Multilingual evaluation of interpretable biomarkers to represent language and speech patterns in Parkinson’s disease. Front Neurol. (2023) 14:1142642. doi: 10.3389/fneur.2023.1142642
31. Pakhomov, SVS, Smith, GE, Chacon, D, Feliciano, Y, Graff-Radford, N, Caselli, R, et al. Computerized analysis of speech and language to identify psycholinguistic correlates of frontotemporal lobar degeneration. Cogn Behav Neurol. (2010) 23:165–77. doi: 10.1097/WNN.0b013e3181c5dde3
32. Agbavor, F, and Liang, H. Predicting dementia from spontaneous speech using large language models. PLOS Digital Health. (2022) 1:e0000168. doi: 10.1371/journal.pdig.0000168
33. Amini, S, Hao, B, Zhang, L, Song, M, Gupta, A, Karjadi, C, et al. Automated detection of mild cognitive impairment and dementia from voice recordings: a natural language processing approach. Alzheimers Dement. (2022) 19:946–55. doi: 10.1002/alz.12721
34. Qiao, Y, Xie, X, Lin, GZ, Zoy, Y, Chen, SD, Ren, RJ, et al. Computer-assisted speech analysis in mild cognitive impairment and Alzheimer's disease: a pilot study from Shanghai. China J Alzheimer’s Dis. (2020) 75:211–21. doi: 10.3233/JAD-191056
35. Themistocleous, C, Eckerstrom, M, and Kokkinakis, D. Identification of mild cognitive impairment from speech in Swedish using deep sequential neural networks. Front Neurol. (2018) 9:975. doi: 10.3389/fneur.2018.00975
36. Ivanova, O, Meilán, JJG, Martínez-Sánchez, F, Martínez-Nicolás, I, Llorente, TE, and González, NC. Discriminating speech traits of Alzheimer’s disease assessed through a corpus of reading task for Spanish language. Comput Speech Lang. (2022) 73:101341. doi: 10.1016/j.csl.2021.101341
37. Martínez-Nicolás, I, Martínez-Sánchez, F, Ivanova, O, and Meilán, JJG. Reading and lexical–semantic retrieval tasks outperforms single task speech analysis in the screening of mild cognitive impairment and Alzheimer’s disease. Sci Rep. (2023) 13:Article 1. doi: 10.1038/s41598-023-36804-y
38. Martínez-Sánchez, F, Meilán, JJG, Vera-Ferrandiz, JA, Carro, J, Pujante-Valverde, IM, Ivanova, O, et al. Speech rhythm alterations in Spanish-speaking individuals with Alzheimer’s disease. Aging Neuropsychol Cognit. (2017) 24:418–34. doi: 10.1080/13825585.2016.1220487
39. Campbell, EL, Docio-Fernandez, L, Jiménez-Raboso, J, and Gacia-Mateo, C. Alzheimer’s dementia detection from audio and text modalities. arXiv. (2020) 2020:4617. doi: 10.48550/arXiv.2008.04617
40. López-de-Ipiña, K, Martinez-de-Lizarduy, U, Calvo, PM, Beitia, B, García-Melero, J, Fernández, E, et al. On the analysis of speech and disfluencies for automatic detection of mild cognitive impairment. Neural Comput Applic. (2018) 32:15761–9. doi: 10.1007/s00521-018-3494-1
41. Lopez-de-Ipina, K, Martinez-de-Lizarduy, U, Calvo, PM, Mekyska, J, Beitia, B, Barroso, N, et al. Advances on automatic speech analysis for early detection of Alzheimer disease: a non-linear multi-task approach. Curr Alzheimer Res. (2018) 15:139–48. doi: 10.2174/1567205014666171120143800
42. Martínez-Nicolás, I, Llorente, TE, Martínez-Sánchez, F, and Meilán, JJG. Ten years of research on automatic voice and speech analysis of people with Alzheimer’s disease and mild cognitive impairment: a systematic review article. Front Psychol. (2021) 12:620251. doi: 10.3389/fpsyg.2021.620251
43. Whiting, PF, Rutjes, AWS, Westwood, ME, Mallett, S, Deeks, JJ, Reitsma, JB, et al. QUADAS-2: a revised tool for the quality assessment of diagnostic accuracy studies. Ann Intern Med. (2011) 155:529–36. doi: 10.7326/0003-4819-155-8-201110180-00009
44. Alzheimer’s Association. 2021 Alzheimer’s disease facts and figures. Alzheimers Dement. (2021) 17:327–406. doi: 10.1002/alz.12328
45. Matthews, KA, Xu, W, Gaglioti, AH, Holt, JB, Croft, JB, Mack, D, et al. Racial and ethnic estimates of Alzheimer’s disease and related dementias in the United States (2015–2060) in adults aged ≥65 years. Alzheimers Dement. (2019) 15:17–24. doi: 10.1016/j.jalz.2018.06.3063
46. Alzheimer’s Association. 2020 Alzheimer’s disease facts and figures. Alzheimers Dement. (2020) 16:391–460. doi: 10.1002/alz.12068
47. Burke, SL, Grudzien, A, Burgess, A, Rodriguez, MJ, Rivera, Y, and Loewenstein, D. The utility of cognitive screeners in the detection of dementia Spectrum disorders in Spanish-speaking populations. J Geriatr Psychiatry Neurol. (2021) 34:102–18. doi: 10.1177/0891988720915513
48. Folstein, MF, Folstein, SE, and McHugh, PR. “Mini-mental state”. A practical method for grading the cognitive state of patients for the clinician. J Psychiatr Res. (1975) 12:189–98. doi: 10.1016/0022-3956(75)90026-6
49. Solomon, PR, Hirschoff, A, Kelly, B, Relin, M, Brush, M, DeVeaux, RD, et al. A 7 minute neurocognitive screening battery highly sensitive to Alzheimer’s disease. Arch Neurol. (1998) 55:349–55. doi: 10.1001/archneur.55.3.349
51. War Department, Adjunct General’s Office. Army individual test battery. Manual of directions and scoring. Washington, DC: War Department, Adjunct General’s Office (1944).
52. Blessed, G, Tomlinson, BE, and Roth, M. The association between quantitative measures of dementia and of senile change in the cerebral grey matter of elderly subjects. Br J Psychiatry. (1968) 114:797–811. doi: 10.1192/bjp.114.512.797
53. Petersen, RC. Mild cognitive impairment as a diagnostic entity. J Internal Med. (2004) 256:183–94. doi: 10.1111/j.1365-2796.2004.01388.x
54. American Psychiatric Association, A. P., & Association, A. P. Diagnostic and statistical manual of mental disorders: DSM-IV, vol. 4. Washington, DC: American psychiatric association (1994).
55. Goodglass, H, Kaplan, E, and Barresi, B. The assessment of aphasia and related disorders. 3rd ed. Philadelphia, PA: Lippincott Williams and Wilkins (2001). 85 p.
56. García-Gutiérrez, F, Marquié, M, Muñoz, N, Alegret, M, Cano, A, De Rojas, I, et al. Harnessing acoustic speech parameters to decipher amyloid status in individuals with mild cognitive impairment. Front Neurosci. (2023) 17:401. doi: 10.3389/fnins.2023.1221401
57. Theodorou, T, Mporas, I, and Fakotakis, N. An overview of automatic audio segmentation. Int J Inf Technol Computer Sci. (2014) 6:1–9. doi: 10.5815/ijitcs.2014.11.01
58. De Cheveigné, A, and Nelken, I. Filters: when, why, and how (not) to use them. Neuron. (2019) 102:280–93. doi: 10.1016/j.neuron.2019.02.039
59. MacCallum, JK, Olszewski, AE, Zhang, Y, and Jiang, JJ. Effects of low-pass filtering on acoustic analysis of voice. J Voice. (2011) 25:15–20. doi: 10.1016/j.jvoice.2009.08.004
60. McFee, B, Raffel, C, Liang, D, McVicar, M, Battenberg, E, and Nieto, O. Librosa: audio and music signal analysis in python. Proceeding of the 14th Python in Science Conference, pp. 18–24. (2015). doi: 10.25080/majora-7b98e3ed-003
61. Honnibal, M, and Montani, I. spaCy 2: natural language understanding with bloom embeddings, convolutional neural networks and incremental parsing. To Appear, No. 7, pp. 411–420. (2017).
62. Pedregosa, F, Varoquaux, G, Gramfort, A, Michel, V, Thirion, B, Grisel, O, et al. Scikit-learn: machine learning in Python. J Mach Learn Res. (2011) 12:2825–30.
63. Zaharia, M, Chen, A, Davidson, A, Ghodsi, A, Hong, SA, Konwinski, A, et al. (2018). Accelerating the machine learning lifecycle with MLflow.
64. Buitinck, L, Louppe, G, Blondel, M, Pedregosa, F, Mueller, A, Grisel, O, et al. API design for machine learning software: experiences from the scikit-learn project. In: Proceedings of the ECML PKDD workshop on languages for data mining and machine learning pp. 108–122. (2013).
65. Refaeilzadeh, P, Tang, L, and Liu, H. Cross-Validation In: L Liu and MT Özsu, editors. Encyclopedia of database systems. New York: Springer (2009). 532–8.
66. Youden, WJ. Index for rating diagnostic tests. Cancer. (1950) 3:32–5. doi: 10.1002/1097-0142(1950)3:1<32::AID-CNCR2820030106>3.0.CO;2-3
67. Sokolova, M, Japkowicz, N, and Szpakowicz, S. Beyond accuracy, F-score and ROC: a family of discriminant measures for performance evaluation In: A Sattar and B Kang, editors. AI 2006: Advances in artificial intelligence. AI 2006. Lecture notes in computer science, vol. 4304. Berlin, Heidelberg: Springer (2006). 1015–21.
68. Khodabakhsh, A, and Demiroglu, C. Analysis of speech-based measures for detecting and monitoring Alzheimer’s disease. Methods Mol Biol. (2015) 1246:159–73. doi: 10.1007/978-1-4939-1985-7_11
69. Fraser, KC, Meltzer, JA, and Rudzicz, F. Linguistic features identify Alzheimer’s disease in narrative speech. J Alzheimers Dis. (2016) 49:407–22. doi: 10.3233/JAD-150520
70. Konig, A, Satt, A, Sorin, A, Hoory, R, Derreumaux, A, David, R, et al. Use of speech analyses within a mobile application for the assessment of cognitive impairment in elderly people. Curr Alzheimer Res. (2015) 15:120–9. doi: 10.2174/1567205014666170829111942
71. Mitchell, AJ. A meta-analysis of the accuracy of the mini-mental state examination in the detection of dementia and mild cognitive impairment. J Psychiatr Res. (2009) 43:411–31. doi: 10.1016/j.jpsychires.2008.04.014
72. Blesa, R, Pujol, M, Aguilar, M, Santacruz, P, Bertran-Serra, I, Hernández, G, et al. Clinical validity of the “mini-mental state” for Spanish speaking communities. Neuropsychologia. (2001) 39:1150–7. doi: 10.1016/s0028-3932(01)00055-0
73. Ciesielska, N, Sokołowski, R, Mazur, E, Podhorecka, M, Polak-Szabela, A, and Kędziora-Kornatowska, K. Is the Montreal cognitive assessment (MoCA) test better suited than the Mini-mental state examination (MMSE) in mild cognitive impairment (MCI) detection among people aged over 60? Meta-analysis. Psychiatr Pol. (2016) 50:1039–52. doi: 10.12740/PP/45368
74. Yamada, Y, Shinkawa, K, Nemoto, M, Nemoto, K, and Arai, T. A mobile application using automatic speech analysis for classifying Alzheimer’s disease and mild cognitive impairment. Comput Speech Lang. (2023) 81:101514. doi: 10.1016/j.csl.2023.101514
75. Fristed, E, Skirrow, C, Meszaros, M, Lenain, R, Meepegama, U, Cappa, S, et al. A remote speech-based AI system to screen for early Alzheimer’s disease via smartphones. Alzheimers Dement (Amst). (2022) 14:e12366. doi: 10.1002/dad2.12366
76. Nieves, ER, Rosenstein, LD, Zaldua, C, Hernando, X, Gabirondo, P, Lacritz, L, et al. H - 52 impact of Spanish dialect on artificial intelligence in assessing expressive language. Arch Clin Neuropsychol. (2023) 38:1536. doi: 10.1093/arclin/acad067.370
77. López-de-Ipiña, K, Alonso, J-B, Travieso, CM, Solé-Casals, J, Egiraun, H, Faundez-Zanuy, M, et al. On the selection of non-invasive methods based on speech analysis oriented to automatic Alzheimer disease diagnosis. Sensors (Basel, Switzerland). (2013) 13:6730–45. doi: 10.3390/s130506730
78. Gosztolya, G, Vincze, V, Tóth, L, Pákáski, M, Kálmán, J, and Hoffmann, I. Identifying mild cognitive impairment and mild Alzheimer’s disease based on spontaneous speech using ASR and linguistic features. Comput Speech Lang. (2019) 53:181–97. doi: 10.1016/j.csl.2018.07.007
79. Nasreen, S, Rohanian, M, Hough, J, and Purver, M. Alzheimer’s dementia recognition from spontaneous speech using disfluency and interactional features. Front Comput Sci. (2021) 3:669. doi: 10.3389/fcomp.2021.640669
80. Arevalo-Rodriguez, I, Smailagic, N, Roqué-Figuls, M, Ciapponi, A, Sanchez-Perez, E, Giannakou, A, et al. Mini-mental state examination (MMSE) for the early detection of dementia in people with mild cognitive impairment (MCI). Cochrane Database of Syst Rev. (2021) 2021:CD010783. doi: 10.1002/14651858.CD010783.pub3
Keywords: digital biomarkers, dementia, mild cognitive impairment, early detection, speech
Citation: Kaser AN, Lacritz LH, Winiarski HR, Gabirondo P, Schaffert J, Coca AJ, Jiménez-Raboso J, Rojo T, Zaldua C, Honorato I, Gallego D, Nieves ER, Rosenstein LD and Cullum CM (2024) A novel speech analysis algorithm to detect cognitive impairment in a Spanish population. Front. Neurol. 15:1342907. doi: 10.3389/fneur.2024.1342907
Edited by:
Catherine C. Price, University of Florida, United StatesReviewed by:
Jeffrey Brooks, University of California, United StatesSamad Amini, Boston University, United States
Huitong Ding, Boston University, United States
Copyright © 2024 Kaser, Lacritz, Winiarski, Gabirondo, Schaffert, Coca, Jiménez-Raboso, Rojo, Zaldua, Honorato, Gallego, Nieves, Rosenstein and Cullum. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: C. Munro Cullum, Munro.Cullum@utsouthwestern.edu