Kathleen C. Fraser

Using linguistic features longitudinally to predict clinical scores for Alzheimer's disease and related dementias

We use a set of 477 lexicosyntactic, acoustic, and semantic features extracted from 393 speech samples in DementiaBank to predict clinical MMSE scores, an indicator of the severity of cognitive decline associated with dementia. We use a bivariate dynamic Bayes net to represent the longitudinal progression of observed linguistic features and MMSE scores over time, and obtain a mean absolute error (MAE) of 3.83 in predicting MMSE, comparable to within-subject interrater standard deviation of 3.9 to 4.8 [1]. When focusing on individuals with more longitudinal samples, we improve MAE to 2.91, which suggests at the importance of longitudinal data collection. Index Terms- Alzheimer’s disease, dementia, Mini-Mental State Examination (MMSE), dynamic Bayes network, feature selection

Speech Recognition in Alzheimer's Disease and in its Assessment.

Narrative, spontaneous speech can provide a valuable source of information about an individual’s cognitive state. Unfortunately, clinical transcription of this type of data is typically done by hand, which is prohibitively time-consuming. In order to automate the entire process, we optimize automatic speech recognition (ASR) for participants with Alzheimer’s disease (AD) in a relatively large clinical database. We extract text features from the resulting transcripts and use these features to identify AD with an SVM classifier. While the accuracy of automatic assessment decreases with increased WER, this is weakly correlated (−0.31). This relative robustness to ASR error is aided by selecting features that are resilient to ASR error.

Sentence segmentation of aphasic speech

Automatic analysis of impaired speech for screening or diagnosis is a growing research field; however there are still many barriers to a fully automated approach. When automatic speech recognition is used to obtain the speech transcripts, sentence boundaries must be inserted before most measures of syntactic complexity can be computed. In this paper, we consider how language impairments can affect segmentation methods, and compare the results of computing syntactic complexity metrics on automatically and manually segmented transcripts. We find that the important boundary indicators and the resulting segmentation accuracy can vary depending on the type of impairment observed, but that results on patient data are generally similar to control data. We also find that a number of syntactic complexity metrics are robust to the types of segmentation errors that are typically made.

Comparison of different feature sets for identification of variants in progressive aphasia

We use computational techniques to extract a large number of different features from the narrative speech of individuals with primary progressive aphasia (PPA). We examine several different types of features, including part-of-speech, complexity, context-free grammar, fluency, psycholinguistic, vocabulary richness, and acoustic, and discuss the circumstances under which they can be extracted. We consider the task of training a machine learning classifier to determine whether a participant is a control, or has the fluent or nonfluent variant of PPA. We first evaluate the individual feature sets on their classification accuracy, then perform an ablation study to determine the optimal combination of feature sets. Finally, we rank the features in four practical scenarios: given audio data only, given unsegmented transcripts only, given segmented transcripts only, and given both audio and segmented transcripts. We find that psycholinguistic features are highly discriminative in most cases, and that acoustic, context-free grammar, and part-of-speech features can also be important in some circumstances.

Using statistical parsing to detect agrammatic aphasia

Agrammatic aphasia is a serious language impairment which can occur after a stroke or traumatic brain injury. We present an automatic method for analyzing aphasic speech using surface level parse features and context-free grammar production rules. Examining these features individually, we show that we can uncover many of the same characteristics of agrammatic language that have been reported in studies using manual analysis. When taken together, these parse features can be used to train a classifier to accurately predict whether or not an individual has aphasia. Furthermore, we find that the parse features can lead to higher classification accuracies than traditional measures of syntactic complexity. Finally, we find that a minimal amount of pre-processing can lead to better results than using either the raw data or highly processed data.

White Matter Disruption and Connected Speech in Non-Fluent and Semantic Variants of Primary Progressive Aphasia

Differential patterns of white matter disruption have recently been reported in the non-fluent (nfvPPA) and semantic (svPPA) variants of primary progressive aphasia (PPA). No single measure is sufficient to distinguish between the PPA variants, but connected speech allows for the quantification of multiple measures. The aim of the present study was to further investigate the white matter correlates associated with connected speech features in PPA. We examined the relationship between white matter metrics and connected speech deficits using an automated analysis of transcriptions of connected speech and diffusion tensor imaging in language-related tracts. Syntactic, lexical, and semantic features were automatically extracted from transcriptions of topic-directed interviews conducted with groups of individuals with nfvPPA or svPPA as well as with a group of healthy controls. A principal component analysis was performed in order to reduce the number of language measures and yielded a five-factor solution. The results indicated that nfvPPA patients differed from healthy controls on a syntactic factor, and svPPA patients differed from controls on two semantic factors. However, the patient groups did not differ on any factor. Moreover, a correlational analysis revealed that the lexical richness factor was significantly correlated with radial diffusivity in the left inferior longitudinal fasciculus, which suggests that semantic deficits in connected speech reflect a disruption of this ventral pathway, and which is largely consistent with the results of previous studies. Using an automated approach for the analysis of connected speech combined with probabilistic tractography, the present findings demonstrate that nfvPPA patients are impaired relative to healthy controls on syntactic measures and have increased radial diffusivity in the left superior longitudinal fasciculus, whereas the svPPA group was impaired on lexico-semantic measures relative to controls and showed increased radial diffusivity in the uncinate and inferior longitudinal fasciculus bilaterally.


Projected Barzilai-Borwein Method with Infeasible Iterates for Nonnegative Least-Squares Image Deblurring

We present a non-monotonic gradient descent algorithm with infeasible iterates for the nonnegatively constrained least-squares deblurring of images. The skewness of the intensity values of the deblurred image is used to establish a criterion for when to enforce the nonnegativity constraints. The approach is observed on several test images to either perform comparably to or to outperform a non-monotonic gradient descent approach that does not use infeasible iterates, as well as the gradient projected conjugate gradients algorithm. Our approach is distinguished from the latter by lower memory requirements, making it suitable for use with large, three-dimensional images common in medical imaging.

Multilingual word embeddings for the assessment of narrative speech in mild cognitive impairment

Abstract We analyze the information content of narrative speech samples from individuals with mild cognitive impairment (MCI), in both English and Swedish, using a combination of supervised and unsupervised learning techniques. We extract information units using topic models trained on word embeddings in monolingual and multilingual spaces, and find that the multilingual approach leads to significantly better classification accuracies than training on the target language alone. In many cases, we find that augmenting the topic model training corpus with additional clinical data from a different language is more effective than training on additional monolingual data from healthy controls. Ultimately we are able to distinguish MCI speakers from healthy older adults with accuracies of up to 63% (English) and 72% (Swedish) on the basis of information content alone. We also compare our method against previous results measuring information content in Alzheimer’s disease, and report an improvement over other topic-modeling approaches. Furthermore, our results support the hypothesis that subtle differences in language can be detected in narrative speech, even at the very early stages of cognitive decline, when scores on screening tools such as the Mini-Mental State Exam are still in the “normal” range.