A carregar...
Projeto de investigação
Sistema de Apoio ao Diagnóstico de Patologias Vocais
Financiador
Autores
Publicações
First version of a support system for the medical diagnosis of pathologies in the larynx
Publication . Fernandes, Joana; Freitas, Diamantino Silva; Teixeira, João Paulo
Voice pathologies are widespread in society. However, the exams are invasive and uncomfortable for the patient, depending on the doctor’s experience doing the evaluation. Classifying and recognizing speech pathologies in a non-invasive way using acoustic analysis saves time for the patient and the specialist while allowing analyzes to be objective and efficient. This work presents a detailed description of an aid system for diagnosing speech pathologies associated with the larynx. The interface displays the parameters that physicians use most to classify subjects: absolute Jitter, relative Jitter, absolute Shimmer, relative Shimmer, Harmonic to Noise Ratio (HNR) and autocorrelation. The parameters used for the classification of the model are also presented (relative Jitter, absolute Jitter, RAP jitter, PPQ5 Jitter, absolute Shimmer, relative Shimmer, shimmer APQ3, shimmer APQ5, fundamental frequency, HNR, autocorrelation, Shannon entropy, entropy logarithmic and subject’s sex), as well as the description of the entire pre-processing of the data (treatment of Outliers using the quartile method, then data normalization and, finally, application of Principal Component Analysis (PCA) to reduce the dimension). The selected classification model is Wide Neural Network, with an accuracy of 98% and AUC of 0.99.
Accuracy Optimization in Speech Pathology Diagnosis with Data Preprocessing Techniques
Publication . Fernandes, Joana Filipa Teixeira; Freitas, Diamantino Rui; Teixeira, João Paulo
Using acoustic analysis to classify and identify speech disorders noninvasively
can reduce waiting times for patients and specialists while also increasing
the accuracy of diagnoses. In order to identify models to use in a vocal disease
diagnosis system, we want to know which models have higher success rates in
distinguishing between healthy and pathological sounds. For this purpose, 708
diseased people spread throughout 19 pathologies, and 194 control people were
used. There are nine sound files per subject, three vowels in three tones, for each
subject. From each sound file, 13 parameters were extracted. For the classification
of healthy/pathological individuals, a variety of classifiers based on Machine
Learning models were used, including decision trees, discriminant analyses, logistic
regression classifiers, naive Bayes classifiers, support vector machines, classifiers
of closely related variables, ensemble classifiers and artificial neural network
classifiers. For each patient, 118 parameters were used initially. The first analysis
aimed to find the best classifier, thus obtaining an accuracy of 81.3% for the
Ensemble Sub-space Discriminant classifier. The second and third analyses aimed
to improve ground accuracy using preprocessingmethodologies. Therefore, in the
second analysis, the PCA technique was used, with an accuracy of 80.2%. The
third analysis combined several outlier treatment models with several data normalizationmodels
and, in general, accuracy improved, obtaining the best accuracy
(82.9%) with the combination of the Greebs model for outliers treatment and the
range model for the normalization of data procedure.
Unidades organizacionais
Descrição
Palavras-chave
Contribuidores
Financiadores
Entidade financiadora
Fundação para a Ciência e a Tecnologia
Programa de financiamento
OE
Número da atribuição
2021.04729.BD