Revolutionizing Early Parkinson's Disease Detection: A Hybrid CNN-LSTM Model

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Sep 7 2023

In a paper published in the journal Scientific Reports, researchers explored early Parkinson's disease (PD) detection through speech analysis by introducing a hybrid model that combined both Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) for dynamic feature extraction from Mel-spectrograms that are derived from speech signals. Their model achieved a remarkable and higher accuracy by surpassing traditional machine learning (ML) approaches in identifying PD, which was demonstrated on the PC-GITA dataset.

Background

PD involves reduced dopamine levels, which impact the person's movement and speech. The early detection of PD is crucial, and ML applied to voice recordings could provide a convenient screening method. Portable devices linked to wearables aid in symptom assessment and medication management. PD presents with symptoms like tremors, rigidity, and speech issues. Levodopa is a key medication that requires precise dosing to balance symptom relief. 

Several recent research studies (Research12-Research33) have focused on detecting and monitoring PD. These studies employ various techniques, including AI, sensor systems, and ML algorithms. They aim to improve early diagnosis and provide valuable insights into the progression of the disease by offering potential benefits for patient care and treatment optimization.

Proposed method

This paper discusses the application of artificial neural networks, Support Vector Machines (SVM), decision trees (CART), XG-Boost, and a proposed hybrid model for PD detection. The utilization of artificial neural networks, which draw inspiration from the human brain, also extends its applications to complex tasks such as document summarization and facial recognition. SVM is a versatile method suitable for both linear and non-linear datasets, and decision trees are fundamental for predictive modeling, while XG-Boost offers efficient tree boosting. 

The proposed hybrid model combines CNN and LSTM to detect PD, which involves data pre-processing, Mel-spectrogram extraction, feature extraction, hyperparameter tuning, and classification. Regularization techniques like L2 regularization and dropout are applied to enhance model performance. This comprehensive overview encompasses various ML techniques and their applications. It culminates in the presentation of the proposed CNN-LSTM hybrid model designed for PD detection. This demonstration underscores the breadth and potential of these approaches, not only in healthcare but also in diverse fields beyond.

Experimental Analysis

This research proposed and implemented a hybrid model in Python to detect PD that included various ML algorithms. The study employed a ten-fold cross-validation approach for model evaluation. The ten-fold cross-validation segments the dataset into ten equal parts for training and testing iteratively. Several performance metrics, including accuracy, recall, precision, and f1-score, were calculated to assess both the proposed hybrid model and existing ML algorithms. These algorithms include Neural Network, SVM, CART, and X-Boost. The hybrid model demonstrated a remarkable accuracy level of 99.49%, surpassing the traditional ML models' performance, which achieved accuracies ranging from 72.69% to 90.81%. These findings suggest the potential utility of the proposed model in clinical care and medical decision-making.

The innovative approach of this study has incorporated dynamic mode decomposition audio samples, mel-spectrograms, and ResNet architectures to extract relevant features for PD diagnosis. The proposed model not only enhanced CNN's performance but also improved PD detection accuracy by combining LSTM and ResNet with CNN. The research underscores the significance of AI-based methodologies in healthcare, particularly for PD diagnosis. It also highlights the potential for future research with larger patient populations. In summary, the proposed hybrid model exhibited exceptional efficiency and performance, offering promising contributions to medical diagnostics.

Contribution of this Paper

Hybrid Model Innovation: The paper introduces a novel hybrid model combining CNN and LSTM networks for PD detection. This innovative approach leverages the strengths of both networks to enhance diagnostic accuracy potentially.

Comprehensive Performance Evaluation: The study conducts a thorough performance assessment of the proposed hybrid model and traditional ML algorithms (Neural Network, SVM, CART, and XG-Boost). It rigorously measures accuracy, recall, precision, and f1-score, providing valuable insights into their effectiveness.

High Accuracy Improvement: The proposed hybrid model demonstrates a significant increase in accuracy compared to conventional ML methods. This improvement has substantial implications for early and precise disease diagnosis, which can lead to better patient outcomes.

Healthcare Implications: The paper underscores the importance of AI-based methodologies in healthcare, particularly for PD diagnosis. It also highlights the potential for future research with larger patient populations by paving the way for further advancements in medical diagnostics.

Conclusion

To sum up, this research introduces a precise hybrid model for PD detection by combining CNN and LSTM alongside other established ML and ensemble methods. The model outperforms existing approaches by achieving a remarkable accuracy of 93.51%. This approach surpasses traditional ML models reliant on static features. Future work will focus on enhancing accuracy through larger datasets and additional patient parameters, addressing the model's time complexity.