A deep learning approach for Parkinson's disease diagnosis based on acoustic signals: LSTM with Bayesian and chaotic optimization

Current diagnostic methods for Parkinson's disease primarily rely on medical observations and the assessment of clinical signs. However, due to the mild nature of early symptoms, their overlap with those of other diseases, and the lack of distinct clinical markers, these early signs are often overlooked or misdiagnosed, thus making the early diagnosis of Parkinson's disease particularly challenging. This study aims to achieve a non-invasive and efficient early diagnosis of Parkinson's disease using deep learning methods based on acoustic signals, while enhancing the model's performance through hyperparameter optimization. All datasets in this study are sourced from publicly available online resources. After extracting multiple acoustic features, the data are merged into a unified dataset for model training. Among the four mainstream neural networks evaluated, long short-term memory (LSTM) demonstrates the best performance, thereby achieving an accuracy of 92.87%, a precision of 95.90%, a recall of 89.57%, and an F1-score of 92.57%. Following hyperparameter optimization using a combined Bayesian and chaotic optimization (BOCO) approach, the LSTM model's accuracy is improved to 94.22%, precision to 97.79%, recall to 90.52%, and F1-score to 93.97%. Consequently, the proposed method in this paper provides an innovative solution for a non-invasive, low-cost, and efficient diagnosis of Parkinson's disease, with broad application prospects.