Convolutional Neural Networks and a Transfer Learning Strategy to Classify Parkinson’s Disease from Speech in Three Different Languages

Vásquez-Correa, Juan Camilo; Arias-Vergara, Tomás; Ríos-Urrego, Cristian David; Schuster, Maria; Rusz, Ján; Orozco-Arroyave, Juan Rafael; Nöth, Elmar

doi:10.1007/978-3-030-33904-3_66

Cited by 17 publications

(7 citation statements)

References 18 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…These models can indeed generalize and adapt to new domains and achieve high performance scores when taking advantage of domain‐specific expertise. Another promising line of research is transfer learning, where complex models can be trained on large and more easily accessible datasets of non‐clinical speech, and then only re‐trained on the smaller clinical datasets; or across clinical datasets (Vásquez‐Correa et al, 2019). Developing such techniques within the field of voice‐based classification of autism is a highly promising venue to foster higher generalizability.…”

Section: Discussionmentioning

confidence: 99%

Vocal markers of autism: Assessing the generalizability of machine learning models

et al. 2022

View full text Add to dashboard Cite

Machine learning (ML) approaches show increasing promise in their ability to identify vocal markers of autism. Nonetheless, it is unclear to what extent such markers generalize to new speech samples collected, for example, using a different speech task or in a different language. In this paper, we systematically assess the generalizability of ML findings across a variety of contexts. We train promising published ML models of vocal markers of autism on novel cross‐linguistic datasets following a rigorous pipeline to minimize overfitting, including cross‐validated training and ensemble models. We test the generalizability of the models by testing them on (i) different participants from the same study, performing the same task; (ii) the same participants, performing a different (but similar) task; (iii) a different study with participants speaking a different language, performing the same type of task. While model performance is similar to previously published findings when trained and tested on data from the same study (out‐of‐sample performance), there is considerable variance between studies. Crucially, the models do not generalize well to different, though similar, tasks and not at all to new languages. The ML pipeline is openly shared. Generalizability of ML models of vocal markers of autism is an issue. We outline three recommendations for strategies researchers could take to be more explicit about generalizability and improve it in future studies. Lay Summary Machine learning approaches promise to be able to identify autism from voice only. These models underestimate how diverse the contexts in which we speak are, how diverse the languages used are and how diverse autistic voices are. Machine learning approaches need to be more careful in defining their limits and generalizability.

show abstract

Section: Discussionmentioning

confidence: 99%

Vocal markers of autism: Assessing the generalizability of machine learning models

et al. 2022

View full text Add to dashboard Cite

show abstract

“…With lower standard deviations of cross-validation results and higher balance in terms of sensitivity and specificity, a classifier based on gradient trees can be considered suitable when dealing with multilingual data. Vasquez-Correa et al [36] also observed a lower difference between sensitivity and specificity and lower standard deviation of cross-validation results when fine-tuning the deep machine learning model. This implies that the model trained on more data has a more balanced classification despite the language differences when choosing an appropriate machine learning approach.…”

Section: Machine Learningmentioning

confidence: 94%

“…Vásquez-Correa et al [36] used convolution neural nets and transfer learning strategy to classify PD in Spanish, German and Czech with MFCC and BBE as input independent variables. Accuracy ranged between 70% and 77%.…”

Section: Introductionmentioning

confidence: 99%

Exploring Digital Speech Biomarkers of Hypokinetic Dysarthria in a Multilingual Cohort

Kovac

Mekyska

Aharonson

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

Motor speech disorders in patients with Parkinson’s disease (PD), collectively referred to as hypokinetic dysarthria, are the early markers of the disease. Acoustic speech features are, therefore, suitable digital biomarkers for the diagnosis and monitoring of this pathological phenomenon. At the same time, it is clear that language plays an essential role in using these features to classify speech successfully into healthy and dysarthric one. This paper focuses on the multilingual analysis of the currently used and newly proposed acoustic speech features and their use in the supportive diagnosis of PD. The goal is to explore digital speech biomarkers of PD, focusing on their language independence and high discrimination power using statistical analysis and machine learning techniques. Thirty-three acoustic features of Czech, American, Israeli, Columbian and Italian PD patients and healthy controls were analyzed using correlation and statistical tests, descriptive statistics, and the XGBoost classifier with posterior explanation by features importances and Shapley values. The features quantifying the prominence of the second formant, monopitch, and the lower number of pauses detected during text reading show the best results, both by statistical analysis and machine learning. Classification accuracies range from 67% to 85%, depending on the language. The paper introduces the concept of language robustness as a quality in which a feature behaves the same, independently of the language. By using this new concept along with other metrics, it proposes several digital speech biomarkers that have the potential to be language-independent with high discrimination power.

show abstract

“…These models can indeed generalize and adapt to new domains and achieve high performance scores when taking advantage of domain-specific expertise. Another promising line of research is transfer learning, where complex models can be trained on large and more easily accessible datasets of non-clinical speech, and then only re-trained on the smaller clinical datasets; or across clinical datasets (Vásquez-Correa et al, 2019). Developing such techniques within the field of voice-based classification of autism is a highly promising venue to foster higher generalizability.…”

Section: Consider the Use Of Multi-dataset ML Techniquesmentioning

confidence: 99%

Vocal markers of autism: assessing the generalizability of machine learning models

Rybner

Jessen

Mortensen

et al. 2021

Preprint

View full text Add to dashboard Cite

Machine learning (ML) approaches show increasing promise in their ability to identify vocal markers of autism. Nonetheless, it is unclear to what extent such markers generalize to new speech samples collected e.g., using a different speech task or in a different language. In this paper, we systematically assess the generalizability of ML findings across a variety of contexts. We train promising published ML models of vocal markers of autism on novel cross-linguistic datasets following a rigorous pipeline to minimize overfitting, including cross-validated training and ensemble models. We test the generalizability of the models by testing them on i) different participants from the same study, performing the same task; ii) the same participants, performing a different (but similar) task; iii) a different study with participants speaking a different language, performing the same type of task.While model performance is similar to previously published findings when trained and tested on data from the same study (out-of-sample performance), there is considerable variance between studies. Crucially, the models do not generalize well to different, though similar, tasks and not at all to new languages. The ML pipeline is openly shared.Generalizability of ML models of vocal markers - and more generally biobehavioral markers - of autism is an issue. We outline three recommendations for strategies researchers could take to be more explicit about generalizability and improve it in future studies.

show abstract

Convolutional Neural Networks and a Transfer Learning Strategy to Classify Parkinson’s Disease from Speech in Three Different Languages

Cited by 17 publications

References 18 publications

Vocal markers of autism: Assessing the generalizability of machine learning models

Vocal markers of autism: Assessing the generalizability of machine learning models

Exploring Digital Speech Biomarkers of Hypokinetic Dysarthria in a Multilingual Cohort

Vocal markers of autism: assessing the generalizability of machine learning models

Contact Info

Product

Resources

About