Otimização metaheurística e aprendizado de máquina para identificação da doença de Parkinson por sinais de voz
Nenhuma Miniatura disponível
Data
2025-01-28
Autores
Garcez, Peter Gleiser
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Federal do Espírito Santo
Resumo
Approximately 220,000 Brazilians have Parkinson’s Disease (PD), which affects 1% to 3%of the world’s population over 65 years, according to WHO estimates. PD causes a continuous and gradual loss of dopamine-producing neurons, a neurotransmitter essential for muscle function performance, especially speech motor control, causing impairment in voice quality. This study aims to implement feature selection and machine learning hyperparameter tuning through optimization metaheuristics to identify PD using features extracted from voice signals. At first, the metaheuristic Adaptive Hybrid-Mutated Differential Evolution (A-HMDE) is applied to select features from the Parkinson’s Disease Classification dataset, consisting of 752 features extracted from 756 voice signal samples. Next, considering the selected features, we tuned the hyperparameter of the Random Forest (RF) and k-Nearest Neighbors (kNN) models, as well as of the Convolutional Neural Network 1D (CNN 1D) model using metaheuristic. A reduction from 752 to 75 features was achieved, representing a selection rate of less than 10%, with an accuracy of 91.63% and a recall of 99.39% obtained by the RF classifier. The results demonstrate the effectiveness of the metaheuristics used for identifying Parkinson’s Disease through voice, and the need to develop datasets with unprocessed vocal signals to explore the performance of convolutional networks operating on raw signals for PD classification
Descrição
Palavras-chave
Palavra-chave , Palavra-chave , Processamento de sinais , Doença de Parkinson , Aprendizado de Máquina