Editorial: Alzheimer's Dementia Recognition through Spontaneous Speech

Abstract

Editorial on the Research Topic Alzheimer's Dementia Recognition through Spontaneous Speech The need for inexpensive, safe, accurate and non-invasive biomarkers for Alzheimer’s disease (AD) has motivated much current research (Mandell and Green, 2011). While diagnosis and evaluation of interventions are still primarily done through clinical assessment, “digital biomarkers” have attracted increasing interest. AI-enabled speech and language analysis has emerged as promising such biomarker for the assessment of disease status (de la Fuente Garcia et al., 2020). While a number of studies have investigated speech and language features for the detection of AD and mild cognitive impairment (Fraser et al., 2016), and proposed various signal processing and machine learning methods for this task (Petti et al., 2020), the field still lacks balanced benchmark data against which different approaches can be systematically compared. This Research Topic addresses this issue by exploring the use of speech characteristics for AD recognition using balanced data and shared tasks, such as those provided by the ADReSS Challenges (Luz et al., 2020, Luz et al., 2021). These tasks have brought together groups working on this active area of research, providing the community with benchmarks for comparison of speech and language approaches to cognitive assessment. Reflecting the multidisciplinary character of the topic, the articles in this collection span three journals: Frontiers of Aging Neuroscience, Frontiers of Computer Science and Frontiers in Psychology. Most papers in this Reseach Topic target two main tasks: AD classification, for distinguishing individuals with AD from healthy controls, and cognitive test score regression, to infer the patient’s Mini Mental Status Examination (MMSE) score (Folstein et al., 1975). Of the twenty papers published in this collection, 14 used the ADReSS dataset (Luz et al., 2020), by itself or in combination with other data. The ADReSS dataset is a curated subset of DementiaBank’s Pitt Corpus, matched for age and gender so as to minimise risk of bias in the prediction tasks. The data consist of audio recordings of picture descriptions elicited from participants using the Cookie Theft picture from the Boston Diagnostic Aphasia Examination (Becker et al., 1994; Goodglass et al., 2001), transcribed and annotated using the CHAT coding system (MacWhinney, 2021). The papers covered a variety of approaches and models. Antonsson et al. aimed to distinguish progressive cognitive decline from stable cognitive impairment using semantic analysis of a discourse task. Support Vector Machine (SVM) models performed best (AUC = 0.93) with both semantic verbal fluency scores and disfluency features from the discourse task. Discourse analysis revealed significantly greater use of unrelated speech in the progressive cognitive decline group compared with the stable group and healthy controls (HC). Clarke et al. examined the impact of five different speech tasks (picture description, conversation, overlearned narrative recall, procedural recall, novel narrative retelling) on classification of 50 participants: 25 HC, 13 mild AD, 12 MCI. Linguistic features (n = 286) were automatically extracted from each task and used to train SVMs. Classification accuracy varied across tasks (62–78% for HC vs AD + MCI, 59–90% for HC vs AD, 50–78% for HC vs MCI) as did which features were most important to the classification. Balagopalan et al. used linguistic and acoustic features derived from ADReSS speech and transcripts. They tuned a pretrained BERT model (Devlin et al., 2018) and compared its features to clinically-interpretable language features. The BERT model outperformed other features and achieved accuracy of 83.33% for AD classification. A ridget regressor with 25 pre-engineered features obtained root mean squared error (RMSE) of 4.56 in MMSE prediction. Chlasta and Wołk used VGGish, a pretrained a Tensorflow model for audio feature extraction and a custom raw waveform based convolutional neural (CNN), DemCNN, to model the acoustic characteristics of AD speech on the ADreSS dataset. DemCNN provided better results than VGGish (Hershey et al., 2017) and achieved an accuracy of 62.5% using only the acoustic information. De Looze et al. combined structural MRI, neuropsychological testing and conversational features to explore temporal characteristics of speech in a collaborative referencing task. They investigated associations with cognitive function and volumetry in brain areas known to be affected by MCI and AD. A linear mixed-effect model was built for data of 32 individuals to assess the predictive power of conversational speech features to classify clinical groups. They found that slower speech and slower turn-taking may provide useful markers for early detection of cognitive decline. Guo et al. emphasized the importance of large normative datasets in training accurate and reliable machine learning models for dementia detection. They incorporated a new corpus of Cookie Theft picture descriptions (HC = 839, NC = 115) from the Wisconsin Longitudinal Study (Herd et al., 2014) to train a BERT model and demonstrated improved performance on the detection task compared with results of the model trained on the ADReSS data alone (82.1% vs 79.8, accuracy, and 92.3 vs 88.3% AUC). Haulcy and Glass investigated the use of i-vectors and x-vectors (Snyder et al., 2018), which are acoustic features originally devised for speaker identification, and linguistic features to tackle AD detection and MMSE prediction. The i-vectors and x-vectors were pre-trained on existing datasets unrelated to AD as well as in-domain data. Several classification and regression models were tested, yielding 85.4% accuracy in AD detection with SVM and Random Forests, and 4.56 RMSE with a gradient boosting regressor. Linguistic and acoustic features were...