1. |
- den Ouden, Dirk B., et al.
(author)
-
Comparison of Automated Methods for Vowel Segmentation and Extraction of Acoustic Variables
- 2018
-
In: Clinical Aphasiology Conference, CAC 2018, Austin, Texas USA..
-
Conference paper (other academic/artistic)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.
|
|
2. |
- Themistocleous, Charalambos, 1980, et al.
(author)
-
A classification study of the variants of Primary Progressive Aphasia using Machine Learning.
- 2018
-
In: Clinical Aphasiology Conference, CAC 2018, Austin, Texas USA.
-
Conference paper (other academic/artistic)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.
|
|
3. |
- Themistocleous, Charalambos, 1980, et al.
(author)
-
Acoustic markers of PPA variants using machine learning.
- 2018
-
In: Frontiers in Human Neuroscience. Conference Abstract: Academy of Aphasia 56th Annual Meeting, October 21-23, 2018, Montreal, Canada. - : Frontiers Media SA. - 1662-5161.
-
Conference paper (other academic/artistic)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.
|
|
4. |
- Themistocleous, Charalambos, 1980, et al.
(author)
-
Assessing Language Disorders using Artificial Intelligence: a Paradigm Shift
- 2023
-
In: arXiv.org. - : arXiv.org.
-
Other publication (other academic/artistic)abstract
- Speech, language, and communication deficits are present in most neurodegenerative syndromes. They enable the early detection, diagnosis, treatment planning, and monitoring of neurocognitive disease progression as part of traditional neurological assessment. Nevertheless, standard speech and language evaluation is time-consuming and resource-intensive for clinicians. We argue that using machine learning methodologies, natural language processing, and modern artificial intelligence (AI) for Language Assessment is an improvement over conventional manual assessment. Using these methodologies, Computational Language Assessment (CLA) accomplishes three goals: (i) provides a neuro-cognitive evaluation of speech, language, and communication in elderly and high-risk individuals for dementia; (ii) facilitates the diagnosis, prognosis, and therapy efficacy in at-risk and language-impaired populations; and (iii) allows easier extensibility to assess patients from a wide range of languages. By employing AI models, CLA may inform neurocognitive theory on the relationship between language symptoms and their neural bases. Finally, it signals a paradigm shift by significantly advancing our ability to optimize the prevention and treatment of elderly individuals with communication disorders, allowing them to age gracefully with social engagement.
|
|