BibTeX

@article{themistocleous-etal-2018-identification-273026,
	title        = {Identification of Mild Cognitive Impairment From Speech in Swedish Using Deep Sequential Neural Networks},
	abstract     = {While people with mild cognitive impairment (MCI) portray noticeably incipient memory difficulty in remembering events and situations along with problems in decision making, planning, and finding their way in familiar environments, detailed neuropsychological assessments also indicate deficits in language performance. To this day, there is no cure for dementia but early-stage treatment can delay the progression of MCI; thus, the development of valid tools for identifying early cognitive changes is of great importance. In this study, we provide an automated machine learning method, using Deep Neural Network Architectures, that aims to identify MCI. Speech materials were obtained using a reading task during evaluation sessions, as part of the Gothenburg MCI research study. Measures of vowel duration, vowel formants (F1 to F5), and fundamental frequency were calculated from speech signals. To learn the acoustic characteristics associated with MCI vs. healthy controls, we have trained and evaluated ten Deep Neural Network Architectures and measured how accurately they can diagnose participants that are unknown to the model. We evaluated the models using two evaluation tasks: a 5-fold crossvalidation and by splitting the data into 90% training and 10% evaluation set. The findings suggest first, that the acoustic features provide significant information for the identification of MCI; second, the best Deep Neural Network Architectures can classify MCI and healthy controls with high classification accuracy (M = 83%); and third, the model has the potential to offer higher accuracy than 84% if trained with more data (cf., SD≈15%). The Deep Neural Network Architecture proposed here constitutes a method that contributes to the early diagnosis of cognitive decline, quantify the progression of the condition, and enable suitable therapeutics.},
	journal      = {Frontiers in Neurology},
	author       = {Themistocleous, Charalambos and Eckerström, Marie and Kokkinakis, Dimitrios},
	year         = {2018},
	volume       = {9},
	pages        = {1--10},
}

@inProceedings{fyndanis-themistocleous-2018-morphosyntactic-271917,
	title        = {Morphosyntactic production in agrammatic aphasia: A cross-linguistic machine learning approach.},
	abstract     = {Introduction
Recent studies on agrammatic aphasia by Fyndanis et al. (2012, 2017) reported evidence against the cross-linguistic validity of unitary accounts of agrammatic morphosyntactic impairment, such as the Distributed Morphology Hypothesis (DMH) (Wang et al., 2014), the two versions of the Interpretable Features’ Impairment Hypothesis (IFIH-1: Fyndanis et al., 2012; IFIH-2: Fyndanis et al., 2018b), and the Tree Pruning Hypothesis (TPH) (Friedmann & Grodzinsky, 1997). However, some of the features/factors emphasized by the accounts above (i.e. involvement of inflectional alternations (DMH), involvement of integration processes (IFIH-1), involvement of both integration processes and inflectional alternations (IFIH-2), position of a morphosyntactic feature/category in the syntactic hierarchy (TPH)) may still play a role in agrammatic morphosyntactic production. These features may act in synergy with other factors in determining the way in which morphosyntactic production is impaired across persons with agrammatic aphasia (PWA) and across languages. Relevant factors may include language-independent and language-specific properties of morphosyntactic categories, as well as subject-specific and task/material-specific variables. The present study addresses which factors determine verb-related morphosyntactic production in PWA and what is their relative importance.

Methods		
We collapsed the datasets of the 24 Greek-, German-, and Italian-speaking PWA underlying Fyndanis et al.’s (2017) study, added the data of two more Greek-speaking PWA, and employed machine learning algorithms to analyze the data. The unified dataset consisted of data on subject-verb agreement, time reference (past reference, future reference), grammatical mood (indicative, subjunctive), and polarity (affirmatives, negatives). All items/conditions were represented as clusters of theoretically motivated features: ±involvement of integration processes, ±involvement of inflectional alternations, ±involvement of both integration processes and inflectional alternations, and low/middle/high position in the syntactic hierarchy. We included 14 subject-specific, category-specific and task/material-specific predictors: Verbal Working Memory (WM), (years of formal) Education, Age, Gender, Mean Length of Utterance in (semi)spontaneous speech (Index 1 of severity of agrammatism), Proportion of Grammatical Sentences in (semi)spontaneous speech (Index 2 of severity of agrammatism), Words per Minute in (semi)spontaneous speech (Index of fluency), Involvement of inflectional alternations, Involvement of integration processes, Involvement of both integration processes and inflectional alternations, Position of a given morphosyntactic category in the syntactic hierarchy (high, middle, low), Item Presentation mode (cross-modal, auditory), Response mode (oral, written), and Language (Greek, German, Italian). Different machine learning models were employed: Random Forest, C5.0 decision tree, RPart, and Support Vector Machine.

Results & Discussion
Random Forest model outperformed all the other models achieving the highest accuracy (0.786). As shown in Figure 1, the best predictors of accuracy on tasks tapping morphosyntactic production were the involvement of both integration processes and inflectional alternations (categories involving both integration processes and inflectional alternations were more impaired than categories involving one or neither of them), verbal WM capacity (the greater the WM capacity, the better the morphosyntactic production), and severity of agrammatism (the more severe the agrammatism, the worse the morphosyntactic production). Results are consistent with IFIH-2 (Fyndanis et al., 2018b) and studies highlighting the role of verbal WM in morphosyntactic production (e.g., Fyndanis et al., 2018a; Kok et al., 2007).},
	booktitle    = {Frontiers in Human Neuroscience. Academy of Aphasia 56th Annual Meeting, Montreal, Canada, 21 Oct - 23 Oct, 2018. },
	author       = {Fyndanis, Valantis and Themistocleous, Charalambos},
	year         = {2018},
}

@inProceedings{fraser-etal-2018-improving-264397,
	title        = {Improving the Sensitivity and Specificity of MCI Screening with Linguistic Information.},
	abstract     = {The Mini-Mental State Exam (MMSE) is a screening tool for cognitive impairment. It has been extensively validated and is widely used, but has been criticized as not being effective in detecting mild cognitive impairment (MCI). In this study, we examine the utility of augmenting MMSE scores with automatically extracted linguistic information from a narrative speech task to better differentiate between individuals with MCI and healthy controls in a Swedish population. We find that with the addition of just four linguistic features, the F score (measuring a trade-off between sensitivity and specificity) is improved from 0.67 to 0.81 in logistic regression classification. These preliminary results suggest that the accuracy of traditional screening tools may be improved through the addition of computerized language analysis.},
	booktitle    = {Proceedings of the LREC workshop: Resources and ProcessIng of linguistic, para-linguistic and extra-linguistic Data from people with various forms of cognitive/psychiatric impairments (RaPID-2). 8th  of May 2018,  Miyazaki, Japan  / Dimitrios  Kokkinakis  (ed.)},
	author       = {Fraser, Kathleen and Lundholm Fors, Kristina and Eckerström, Marie and Themistocleous, Charalambos and Kokkinakis, Dimitrios},
	year         = {2018},
	ISBN         = {979-10-95546-26-9},
}

@inProceedings{themistocleous-etal-2018-acoustic-271915,
	title        = {Acoustic markers of PPA variants using machine learning.},
	abstract     = {Introduction. Speakers’ acoustic profile carries significant linguistic and non-linguistic information. Employed in clinical practice, it can provide behavioral markers for a quick assessment of primary progressive aphasia (PPA). PPA is a complex language syndrome where different speech and language properties such as prosody, lexical retrieval, and motor speech functioning may be affected.  It is classified into three main variants: the nonfluent (nfvPPA), semantic (svPPA), and logopenic (lvPPA). Primary progressive apraxia of speech (PPAOS) is also distinguished (Duffy et al. 2017) but may fall into the category of nfvPPA (Gorno-Tempini et al. 2011). The present study aims to determine the contribution of the acoustic properties of vowels, prosody, and voice quality in the classification of PPA variants by using machine learning models.

Methods. Oral samples from picture description tasks of 50 individuals with PPA (lvPPA:17, svPPA:14, nfvPPA:11, PPAOS:8) were automatically transcribed and segmented into vowels and consonants using the new acoustic analysis platform THEMIS. From the segmented vowels, we measured: i. Vowel formants (F1…F5) (den Ouden, et al. 2017); ii. vowel duration (Duffy, et al., 2017); iii. Mean fundamental frequency (F0), min F0 and max F0 (Hillis, 2014); iv. Pause duration (Mack et al. 2015), and v. H1–H2, H1–A1, H1–A2, H1–A3 measures of voice quality. We compared three machine learning models: support vector machines (SVM) (Cortes and Vapnik, 1995), random forests (RF) (Breiman, 2001), and decision trees (DT) (Hastie et al. 2009) in an one-against all strategy, where each variant was tested against all others. We run all models with a 3-fold group-cross-validation to ensure that the speakers in the training and evaluation sets are different. The models were implemented in Python (Pedregosa et al. 2011).

Results. We report the mean cross-validated accuracy of the best performing model that resulted from model comparison: i. RF model provided the highest classification accuracy for nfvPPA [Mean 82%, SD: 9%], ii. SVM had the highest accuracy for svPPA [Mean 66%, SD: 8%], iii. RF had the highest accuracy for lvPPA [Mean 57%, SD: 15%] and iv. RF provided the highest classification accuracy for PPAOS [Mean 80%, SD: 8%] (Figure 1). In all models, pause duration and F0 measures were ranked higher than most other features (Figure 2).

Discussion. This study employed an innovative method for the classification of PPA variants, using an automated speech transcription, segmentation, feature extraction and modeling. Using just acoustic features the best model classified nfvPP, svPPA, and PPAOS with high accuracy. However, acoustic features alone could not classify lvPPA with such high accuracy. More linguistic markers might be needed for a more accurate classification of lvPPA. Furthermore, we showed that prosody, which is measured by fundamental frequency and pause duration, contributes more than any other factor to the classification of PPA variants as alluded in previous research by our group and others (Hillis 2014, Patel et al. 2018, Mack 2015). Finally, the findings demonstrate the potential benefit of using machine learning models in clinical practice for the subtyping of PPA variants.},
	booktitle    = {Frontiers in Human Neuroscience. Conference Abstract: Academy of Aphasia 56th Annual Meeting, October 21-23, 2018, Montreal, Canada},
	author       = {Themistocleous, Charalambos and Ficek, Bronte and Webster , Kimberly and Wendt, Haley  and Hillis, Argye and  Den Ouden , Dirk-Bart  and Tsapkini, Kyrana},
	year         = {2018},
}

@inProceedings{neofytou-etal-2018-understanding-271916,
	title        = {Understanding and classifying the different variants of Primary Progressive Aphasia based on spelling performance},
	abstract     = {Introduction: Previous findings suggest differences in the written spelling performance between the three variants of Primary Progressive Aphasia (PPA) - semantic (svPPA), logopenic (lvPPA) and non-fluent (nfvPPA) (Shim et al., 2012; Sepelyak et al., 2011). However, no attempts have been made to systematically distinguish the three variants in terms of their spelling performance. The challenges of classification are considerable and given the ease of administering a spelling test, we aimed to determine to what extent a spelling task can provide accurate classification of the  PPA variants.
Method: Thirty-three participants with PPA were included - 14 lvPPAs, 11 nfvPPAs and 8 svPPAs – originally classified using the neuropsychological and spoken language criteria defined by Gorno-Tempini et al. (2011). Data were collected prior to spelling treatment, using a spelling to dictation task with both real-words and pseudowords (92-138 items/per participant), scored for each grapheme (i.e., letter) and analyzed for each participant individually using generalized linear mixed effects models (GLMEM) for real-words and pseudowords separately. The variables of interest for both real-words and pseudowords were word length, phoneme-grapheme conversion probability and grapheme position. The real-word models also included frequency, imageability, and the orthographic and phonological neighborhood density of the target words. 
The coefficients from the output of the GLMEMs, together with 3 additional variables – verb/noun  and pseudoword/word accuracy differences from the spelling task, and language impairment severity according to FTD-CDR (Knopman, 2008) - were used as predictors in a Random Forests (RFs) model implemented in Python, to identify the variables that contribute the most in distinguishing the three variants. Then, the three most significant predictors identified with RFs were used in multinomial models implemented in R to classify the PPA variants. The model was trained on a training set of all participants minus one (i.e. the left-out participant) and evaluated on the left-out participant, known as Leave-One-Out cross-validation. This process was repeated 33 times to evaluate all participants. 
Results: The three most significant predictors of the RFs analysis were: (1) grapheme position in real-words, (2) pseudoword/word accuracy difference, and (3) length of real-words (Figure 1). The overall accuracy of the multinomial models with these three predictors only was 67%: lvPPA=71%, nfvPPA=64% and svPPA=63%. When severely impaired cases (language severity =3 in Knopman et al., 2008; FTD-CDR criteria) were excluded (giving a new dataset of 22 participants), the overall accuracy increased to 91%: lvPPA=90%, nfvPPA=86% and svPPA=100%.
Discussion: Our study provides evidence of the value of considering spelling performance in understanding and classifying the different variants of PPA. The results suggest that lexical status, word length and grapheme position are useful parameters for classification, which index key components of the cognitive architecture of spelling (Rapp, 2002). Also, the finding that prediction accuracy increased when more severe cases were excluded supports  previous findings (Mesulam et al., 2012), as severity increases variants become less differentiated and classification is more difficult. In sum, a relatively short, easy-to-administer spelling test, provides useful information for PPA variant classification and can  potentially be used as a clinical tool.},
	booktitle    = {Frontiers in Human Neuroscience},
	author       = {Neofytou, Kyriaci and Themistocleous, Charalambos and Wiley,  Robert  and Tsapkini, Kyrana and Rapp, Brenta},
	year         = {2018},
}

@inProceedings{themistocleous-etal-2018-effects-270215,
	title        = {Effects of Mild Cognitive Impairment on vowel duration
},
	abstract     = {Mild cognitive impairment (MCI) is a neurological condition, which is characterized
by a noticeable decline of cognitive abilities, including communicative and linguistic
skills. In this study, we have measured the duration of vowels produced in a reading
task by 55 speakers— 30 healthy controls and 25 MCI—. The main results showed
that MCI speakers differed significantly from HC in vowel duration as MCI speakers
produced overall longer vowels. Also, we found that gender effects on vowel duration
were different in MCI and HC. One significant aspect of this finding is that they highlight
the contribution of vowel acoustic features as markers of MCI.},
	booktitle    = {Proceedings of the 9th Tutorial & Research Workshop on Experimental Linguistics, 28 - 30 August 2018, Paris, France},
	editor       = {Antonis Botinis},
	author       = {Themistocleous, Charalambos and Kokkinakis, Dimitrios and Eckerström, Marie and Fraser, Kathleen and Lundholm Fors, Kristina},
	year         = {2018},
	ISBN         = {978-960-466-162-6 },
}

@inProceedings{themistocleous-kokkinakis-2018-themis-265112,
	title        = {THEMIS-SV: Automatic classification of language disorders from speech signals},
	abstract     = {Background and Aims:
Brain injuries resulting from stroke can affect the production of speech resulting in different types of language impairments, such as aphasia.
Studying these productions manually is an extremely cumbersome and time consuming process. 
The aim of this paper is to present THEMIS-SV: a system that enables the automatic transcription of speech signals and the segmentation of vowels and consonants in Swedish.

Method:
The input of the system are recordings of speech. The system processes the recordings and returns an output with three tiers: the utterance tier, the word tier, and the vowels/consonants tier. 

Results:
The output of the system is a fast and reliable transcription and segmentation of speech, which is very close to transcriptions and segmentations performed manually. 
The automatic segmentation of speech enables targeted acoustic measurements, such as measurements of consonant spectra, formant frequencies of vowels, fundamental frequency, pauses, speech rate, etc. and other acoustic measurements that have been known to differentiate between the different types of language disorders.

Conclusion:
The method proposed here can be employed for the analysis of speech of individuals with post-stroke aphasia and other language disorders and constitutes a promising step towards a fully automated differential diagnostic tool for language disorders. },
	booktitle    = {Abstracts of the 4th European Stroke Organisation Conference ​(ESOC 2018). Gothenburg, Sweden, 16-18 May, 2018. },
	author       = {Themistocleous, Charalambos and Kokkinakis, Dimitrios},
	year         = {2018},
}

@inProceedings{kokkinakis-etal-2018-textforskning-265113,
	title        = {Kan textforskning bidra till tidigare och säkrare demensdiagnostik?},
	abstract     = {Tidigare forskning har visat att subtila språkstörningar kan finnas vid de tidigaste förstadierna till demens, flera år innan en klinisk diagnos kan ställas. Inom ramen för projektet ”Språkliga och extra-lingvistiska parametrar för tidig upptäckt av kognitiv svikt” (finansierat av Riksbankens Jubileumsutlysning, 2016-19) undersöker vi med hjälp av språkteknologi och språkanalysstudier hur dessa språkstörningar yttrar sig. Kan språkteknologi användas för att upptäcka dessa tidiga språkrelaterade symtom och därmed bidra med nyanserad, komplementär och användbar kunskap? Kan användning av språkteknologi särskilja personer med de allra tidigaste kognitiva avvikelserna från personer med mer godartad, åldersrelaterad kognitiv svikt? Vilka språkliga förmågor drabbas? Hur yttrar sig dessa förändringar och vilka slags empiriska material finns att tillgå? Dessa är några av de frågor vi söker svar på. Vi gör inspelningar som vi analyserar för att kunna ta fram ny kunskap om subtila språkliga kännetecken som kan föregå demensutveckling. Denna kunskap kan användas för att eventuellt kunna förutsäga vilka individer som befinner sig i riskzonen för att utveckla demens, och kan vara användbar som komplementerande beslutsunderlag till domänexperter. Vi utvinner, analyserar och undersöker om det finns samband mellan olika språkrelaterade parametrar från spontan talinteraktion, transkriptioner men även ögonrörelser och neuropsykologiska tester från personer med subjektiv eller lindrig kognitiv nedsättning och friska kontrollpersoner. Många gånger är det svårt att avgöra huruvida lindriga kognitiva symtom är en del av det normala åldrandet eller början på en neurodegenerativ process. Vi förväntar oss inte heller att varje enskild person med kognitiv nedsättning kommer att uttrycka sig eller läsa på samma sätt utan snarare att dessa personer tidigt i sjukdomsförloppet kommer att börja uppvisa olika slags avvikande läsmönster, eller göra fonologiska, lexikala, syntaktiska eller semantiska fel. I studien utvecklar vi verktyg för att automatiskt hitta dessa avvikelser, och målet är att detta sedan ska kunna användas som komplement till tidig diagnostik samt som prognostiskt eller screeningverktyg. Deltagarna i vår studie har rekryterats från en pågående longitudinell studie, ”Demens i Tidigt Skede”, (eng. ”The Gothenburg MCI study”) på Minnesmottagningen i Göteborg, och vårt projekt har godkänts av den lokala etiknämnden. Alla deltagare i studien (kontrollgruppen [HC], personer med subjektiv kognitiv nedsättning [SCI] och personer med mild kognitiv nedsättning [MCI]) har genomgått baslinjeundersökning och gett informerat skriftligt samtycke (demografisk information finns i tabell 1). Vårt projekt är f.n. pågående och vi kommer presentera resultat baserade på inspelningstillfälle nr ett (aug. 2016-mars 2017). En ny inspelningsomgång, med samma deltagare, började i februari 2018 och förväntas vara avslutat i december 2018. Under presentationen kommer vi ge exempel på olika tal-, text- och ögonrörelseanalyser vi har genomfört och diskutera metodval och resultat baserade på studiens första fas. Vi kommer vidare ge en kort inblick i den nya, pågående inspelningsomgången och de nya testmoment vi använder. Vi vill med vårt arbete visa hur språkteknologisk analys kan bidra till att utöka vår kunskap inom området så att den kan vara användbar för tidig diagnostik och optimal omvårdnad. Enligt Socialstyrelsen (2017) finns det i Sverige över 160 000 personer med någon demenssjukdom. Våra resultat kan ha en betydelse för vårdpersonal som snabbare vill diagnostisera och identifiera individer med olika former av kognitiv funktionsnedsättning innan allvarliga symtom blir påtagliga. Utvecklingsmöjligheterna är många: nya eller förbättrade kognitiva screeningtester som skulle kunna användas inom primär- och specialistvården, samt utveckling och tillämpning av insatser som kan påverka beteendemönster och träna upp individens kommunikativa förmåga, kan på sikt leda till positiva konsekvenser som minskade vårdköer samt effektivare behandling avseende kostnader och behandlingsutfall.},
	booktitle    = {Forum för textforskning 13 , Lund 7 – 8 juni 2018},
	author       = {Kokkinakis, Dimitrios and Lundholm Fors, Kristina and Eckerström, Marie and Themistocleous, Charalambos},
	year         = {2018},
}

@inProceedings{themistocleous-etal-2018-classification-268340,
	title        = {A classification study of the variants of Primary Progressive Aphasia using Machine Learning.},
	abstract     = {Introduction: Primary Progressive Aphasia (PPA) is a neurodegenerative syndrome in which linguistic abilities become gradually impaired. There are three primary variants of PPA: the non-fluent agrammatic PPA, the fluent type semantic PPA, and the logopenic PPA, which is also considered an atypical form of Alzheimer’s disease (Mesulam et al., 1982; Gorno-Tempini et al., 2011). Along with the three main variants, a fourth variant has been proposed, a non-fluent apraxia of speech (AOS), though this is currently the subject of an open debate (e.g., Duffy et al., 2017; Henry et al., 2013). According to sophisticated criteria established a few years ago, PPA subtyping for a given patient presented in clinic requires clinical, neuropsychological, and imaging information (Gorno-Tempini et al., 2011). Nevertheless, quantifying the decline of linguistic abilities and subtyping its variants manually is both hard and laborious, so there is a great demand for algorithms that subtype a given patient automatically. Picture description samples of connected speech and random forests techniques have been used for this purpose (de Aguiar et al., 2017; Wilson et al., 2010, Fraser et al. 2013, 2014). In the present study, we compared existing models and we propose a new one.
Aims: In this study, we provide an automated classification model of the four PPA variants trained on known morphological and acoustic predictors and on predictors related to the clinical and linguistic profile of individuals with PPA (e.g., Mack et al., 2015; Gorno-Tempini et al., 2011; Wilson et al., 2010). 
Method: Speech materials for this study come from the Transcranial Direct Current Stimulation for Primary Progressive Aphasia study at Johns Hopkins University. Twenty-six individuals with PPA (Mean(SD) age = 68.6 (7.8) years, Mean(SD) education = 16.1 (2.9) years) participated in this study. PPA participants were diagnosed based on the established consensus criteria (Gorno-Tempini et al., 2011) based on imaging, clinical, and neuropsychological examination by trained neurologists. Individuals with PPA included non-fluent AOS (N=5), non fluent (N=7), logopenic (N=8), and semantic (N=6) variants. 
Recordings of the Cookie Theft picture description from the Boston Diagnostic Aphasia Examination (BDAE) were computationally analyzed. All speech productions were automatically transcribed and segmented using an end-to-end speech-to-transcription platform. From the speech signals, we measured morphological and acoustic predictors, including vowel formants F1 ... F3, measured at 15%, 50%, and 75% of vowel’s duration, vowel duration, fundamental frequency, and pause duration. The analysis and the statistics were conducted using Python and R programming languages (R Core Team, 2017; Rossum, 1995). 
Three different machine learning algorithms: C5.0 decision trees, Classification and Regression Trees (CART) and random forests were trained on the predictors (Breiman, 2001; Quinlan, 1993; Hastie et al., 2009). All models were trained on the 80% of the speakers (training set), with 3-fold cross-validation. All predictor variables were centered and scaled. C5.0 was trained with winnowing and without winnowing. (Winnowing facilitates the automatic pre-selection of the predictors that are used in the decision tree.) After the training we evaluated the trained models on the unknown dataset, namely the 20% of the speakers (evaluation set).
Results: C5.0 provided 86% (95% CI[81, 88], kappa = 0.76) and Random Forests 85% (95% CI[81, 88], kappa = 0.76) classification accuracy on the test data; CART provided the lowest overall classification accuracy. Overall, C5.0 outperformed both the random forests and CART, with high classification accuracy on unknown data. Non-fluent AOS was correctly predicted by both C5.0 and random forests.
Discussion: C5.0 classification model provides support for the known predictors employed in the literature. Also, it provides initial support for the distinct properties of the non-fluent AOS variant and corroborate research on classification of AOS using acoustic properties especially those related to vowel production (Den Ouden et al. 2017). However, given the low number of participants employed in this study, further research is required, with a largest number of participants. Nevertheless, the proposed methods employed here constitute a promising step towards a computational differential diagnostic tool of PPA that is easy to use, quick and accurate. },
	booktitle    = {Clinical Aphasiology Conference, CAC 2018, Austin, Texas USA},
	author       = {Themistocleous, Charalambos and Ficek, Bronte and Webster, Kimberly and Wendt, Haley and Hillis, Argye E. and den Ouden , Dirk Bart and Tsapkini, Kyrana},
	year         = {2018},
}

@inProceedings{denouden-etal-2018-comparison-268339,
	title        = {Comparison of Automated Methods for Vowel Segmentation and Extraction of Acoustic Variables},
	abstract     = {Introduction: Primary Progressive Aphasia (PPA) is a neurodegenerative syndrome in which linguistic abilities become gradually impaired. There are three primary variants of PPA: the non-fluent agrammatic PPA, the fluent type semantic PPA, and the logopenic PPA, which is also considered an atypical form of Alzheimer’s disease (Mesulam et al., 1982; Gorno-Tempini et al., 2011). Along with the three main variants, a fourth variant has been proposed, a non-fluent apraxia of speech (AOS), though this is currently the subject of an open debate (e.g., Duffy et al., 2017; Henry et al., 2013). According to sophisticated criteria established a few years ago, PPA subtyping for a given patient presented in clinic requires clinical, neuropsychological, and imaging information (Gorno-Tempini et al., 2011). Nevertheless, quantifying the decline of linguistic abilities and subtyping the variants of PPA manually is both hard and laborious, so there is great demand for algorithms that subtype a given patient automatically. Picture description samples of connected speech and random forests techniques have been used for this purpose (de Aguiar et al., 2017; Wilson et al., 2010, Fraser et al. 2013, 2014). In the present study, we compared existing models and we propose a new one.
Aims: In this study, we provide an automated classification model of PPA variants trained on known morphological and acoustic predictors and on predictors related to the clinical and linguistic profile of individuals with PPA (e.g., Mack et al., 2015; Gorno-Tempini et al., 2011; Wilson et al., 2010). 
Method: Speech materials for this study come from the Transcranial Direct Current Stimulation for Primary Progressive Aphasia study at Johns Hopkins University. Twenty-six individuals with PPA (Mean(SD) age = 68.6 (7.8) years, Mean(SD) education = 16.1 (2.9) years) participated in this study. PPA participants were diagnosed based on the established consensus criteria (Gorno-Tempini et al., 2011), i.e., imaging, clinical, and neuropsychological examination by trained neurologists. Individuals with PPA included non-fluent with AOS (N=5), non fluent without AOS (N=7), logopenic (N=8), and semantic (N=6) variants. 
Recordings of the Cookie Theft picture description from the Boston Diagnostic Aphasia Examination (BDAE) were computationally analyzed. All speech productions were automatically transcribed and segmented using an end-to-end speech-to-transcription platform. From the speech signals, we measured morphological and acoustic predictors, including vowel formants F1 ... F3, measured at 15%, 50%, and 75% of vowel’s duration, vowel duration, fundamental frequency, and pause duration. The analysis and the statistics were conducted using Python and R programming languages (R Core Team, 2017; Rossum, 1995). 
Three different machine learning algorithms: C5.0 decision trees, Classification and Regression Trees (CART) and random forests were trained on the predictors (Breiman, 2001; Quinlan, 1993; Hastie et al., 2009). All models were trained on the 80% of the speakers (training set), with 3-fold cross-validation. All predictor variables were centered and scaled. C5.0 was trained with winnowing and without winnowing. (Winnowing facilitates the automatic pre-selection of the predictors that are used in the decision tree.) After the training we evaluated the trained models on the unknown dataset, namely the 20% of the speakers (evaluation set).
Results: C5.0 provided 86% (95% CI[81, 88], kappa = 0.76) and Random Forests 85% (95% CI[81, 88], kappa = 0.76) classification accuracy on the test data; CART provided the lowest overall classification accuracy. Overall, C5.0 outperformed both the random forests and CART, with high classification accuracy on unknown data. Non-fluent PPA with AOS was correctly predicted by both C5.0 and random forests.
Discussion: The C5.0 classification model provides support for the known predictors employed in the literature. Also, it provides some objective ways to distinguish the presence of AOS in PPA and corroborate research on classification of AOS using acoustic properties especially those related to vowel production (Den Ouden et al. 2017). However, given the low number of participants employed in this study, further research is required, with a larger number of participants. Nevertheless, the proposed methods employed here constitute a promising step towards a computational differential diagnostic tool of PPA that is easy to use, quick and accurate. },
	booktitle    = {Clinical Aphasiology Conference, CAC 2018, Austin, Texas USA.},
	author       = {den Ouden, Dirk B. and Hutchinson, Angelica and Tsapkini, Kyrana and Themistocleous, Charalambos},
	year         = {2018},
}

@inProceedings{angelopoulou-etal-2018-pause-268338,
	title        = {Pause patterns and speech errors in stroke patients with aphasia: cross-linguistic evidence from narrative speech.},
	booktitle    = {Clinical Aphasiology Conference, CAC 2018, Austin, Texas USA.},
	author       = {Angelopoulou, Georgia and Kiran, Swathi and  Kasselimis, Dimitrios and Varkanitsa, Maria and Meier, Erin and Yue, Pan and Tsolakopoulos, Dimitrios and Themistocleous, Charalambos and Vassilopoulou, Sofia and Korompoki , Eleni and Tountopoulou, Argyro and Papageorgiou, Georgios and Goutsos, Dionysis, and Evdokimidis, Ioannis and Potagas, Constantin},
	year         = {2018},
}