Mecanismos de atenção multi-head em modelos recorrentes e análise acústica da voz para auxílio ao diagnóstico da doença de Parkinson.

Abstract

Parkinson’s disease (PD) is a degenerative neurological disorder characterized by reduced dopamine production, resulting in motor and non-motor symptoms. Speech alterations are promising biomarkers for early diagnosis of the disease, motivating this study to investigate the use of voice and different Recurrent Neural Network (RNN) architectures, with and without multi-head attention mechanisms, in the automatic classification of Parkinson’s disease. The RNNs use Mel Frequency Cepstral Coefficients (MFCC) as input, with different window sizes, enabling the capture of acoustic speech patterns. Statistical analyses confirmed that the use of multi-head attention significantly improves the performance of Bidirectional Long Short-Term Memory (BiLSTM) models, especially with 1,200-frame windows (F1-score = 0.7468; ROC-AUC = 0.7848). However, in Bidirectional Gated Recurrent Unit (BiGRU) models, the inclusion of multi-head attention did not yield significant improvement, indicating that the impact of attention depends on the architecture. Cross-validation assessments (5-fold) confirmed the robustness of the results, with the BiLSTM model using multi-head attention (2 heads, 1,200 frames) achieving performance of F1-score = 0.870, despite the influence of data heterogeneity. These findings reinforce the feasibility of using speech signals and recurrent neural networks with attention for the early detection of Parkinson’s disease, highlighting the importance of balancing window size and model complexity.