|
|
|
| e-Pub |
Previous | 
Home
Section: New Results
Speech-Based Analysis for older people with dementia
Participants : Alexandra König, Philippe Robert, Nicklas Linz, Johannes Tröger, Jan Alexandersson.
Keywords: Alzheimer's disease, Dementia, Mild cognitive impairment, Neuropsychology , Assessment, Semantic verbal fluency, Speech recognition, Speech processing, Machine learning
Fully Automatic Speech-Based Analysis of the Semantic Verbal Fluency Task:
Semantic verbal fluency (SVF) tests are routinely used in screening for mild cognitive impairment (MCI). In this task, participants name as many items as possible of a semantic category under a time constraint. Clinicians measure task performance manually by summing the number of correct words and errors. More fine-grained variables add valuable information to clinical assessment, but are time-consuming. Therefore, the aim of this study is to investigate whether automatic analysis of the SVF could provide measures as accurate as the manual ones and thus, support qualitative screening of neurocognitive impairment.
Methods: SVF data were collected from 95 older people with MCI (n = 47), Alzheimer’s or related dementias (ADRD; n = 24), and healthy controls (HC; n = 24). All data were annotated manually and automatically with clusters and switches. The obtained metrics were validated using a classifier to distinguish HC, MCI, and ADRD.
Results: Automatically extracted clusters and switches were highly correlated (r = 0.9) with manually established values, and performed as well on the classification task, separating HC from persons with ADRD (area under curve [AUC] = 0.939) and MCI (AUC = 0.758).
Conclusion: The results show that it is possible to automate fine-grained analyses of SVF data for the assessment of cognitive decline [70].
Language Modelling in the Clinical Semantic Verbal Fluency Task:
We employed language modelling (LM) as a natural technique to model production in this task. Comparing different LMs, we show that perplexity of a person's SVF production predicts dementia well (F1 = 0.83). Demented patients show significantly lower perplexity, thus are more predictable. Persons in advanced stages of dementia differ in predictability of word choice and production strategy - people in early stages differ only in predictability of production strategy (Linz et al., 2018a).
Telephone-based Dementia Screening I: Automated Semantic Verbal Fluency Assessment:
Despite encouraging results, there are still two main issues in leveraging pervasive sensing technologies for automatic dementia screening: significant hardware costs or installation efforts and the challenge of an effective pattern recognition. Conversely, automatic speech recognition (ASR) and speech analysis have reached sufficient maturity and allow for low-tech remote telephone-based screening scenarios. Therefore, we examine the technological feasibility of automatically assessing a neuropsychological test—Semantic Verbal Fluency (SVF)–via a telephone-based solution. We investigate its suitability for inclusion into an automated dementia frontline screening and global risk assessment, based on concise telephone-sampled speech, ASR and machine learning classification. Results are encouraging showing an area under the curve (AUC) of 0.85. We observe a relatively low word error rate of 33% despite phone-quality speech samples and a mean age of 77 years of the participants. The automated classification pipeline performs equally well compared to the classifier trained on manual transcriptions of the same speech data. Our results indicate SVF as a prime candidate for inclusion into an automated telephone-screening system [50].
Using Acoustic Markers extracted from Free Emotional Speech:
Apathy is a frequent neuropsychiatric syndrome in people with dementia. It leads to diminished motivation for physical, cognitive and emotional activity. Apathy is highly underdiagnosed since its criteria have been only recently established and rely heavily on the subjective evaluation of human observers. We analyzed speech samples from demented people with and without apathy. Speech was provoked by asking patients two emotional questions. Acoustic features were extracted and used in a classification task. The resulting models show performances of AUC = 0.71 and AUC = 0.63. This is a decent first step into the direction of automatic detection of apathy from speech. Usefulness of stimuli to elicit free speech is found to depend on patients' gender [46].
Using Automatic Speech Analysis:
Apathy is present in several psychiatric and neurological conditions and found to have a severe negative effect on patients' life. In older people, it can be a predictor of increased dementia risk. Current assessment methods seem insufficiently objective and sensitive, thus new diagnostic tools and broad-scale screening technologies are needed. This study is the first of its kind aiming to investigate whether automatic speech analysis could be used for characterization and detection of apathy.
Methods: A group of apathetic and non-apathetic patients (n = 60) was recorded while performing two short narrative speech tasks. Paralinguistic markers relating to prosodic, formant, source and temporal qualities of speech were automatically extracted, examined between the groups and compared to baseline assessments. Machine learning experiments were carried out to validate the diagnosis power of extracted markers.
Results: Correlations between apathy sub-scales and features revealed a relation between temporal aspects of speech and the subdomains of reduction in interest and initiative, as well as between prosody features and the affective domain. Group differences were found to vary for males and females, depending on the task. Differences in temporal aspects of speech were found to be the most consistent difference between apathetic and non-apathetic patients. Machine learning models trained on speech features achieved top performances of AUC = 0.88 for males and AUC = 0.77 for females (article under review).
An additional study in this context analyses transcripts of responses to emotional questions (positive and negative) for sentiment using a French emotion dictionary (FEEL) and for psycholinguistic properties (LIWC). Significant reductions in the number of words, the magnitude of sentiment, the overall sentiment and the range between sentiment in the positive and negative questions are found for the apathetic population. This effect is consistent between the positive and the negative stories. When training machine learning classifiers to detect apathy based on these features, the best model showed an AUC of 0.874 using only sentiment features. LIWC features mostly showed no predictive power. When ASR technology was introduced to automatically create transcripts, the performance of predictive models dropped slightly to AUC = 0.864. ASR errors were consistent over all categories of sentiment words. These results highlight the potential of computational linguistic analysis in screening for apathy (article under review).