Abstract
The risk of Parkinson's disease (PD) is extremely serious, and PD speech recognition is an effective method of diagnosis nowadays. However, due to the influence of the disease stage, corpus, and other factors on data collection, the ability of every samples within one subject to reflect the status of PD vary. No samples are useless totally, and not samples are 100% perfect. This characteristic means that it is not suitable just to remove some samples or keep some samples. It is necessary to consider the sample transformation for obtaining high quality new samples. Unfortunately, existing PD speech recognition methods focus mainly on feature learning and classifier design rather than sample learning, and few methods consider the sample transformation. To solve the problem above, a PD speech sample transformation algorithm based on multitype reconstruction operators is proposed in this paper. The algorithm is divided into four major steps. Three types of reconstruction operators are designed in the algorithm: types A, B and C. Concerning the type A operator, the original dataset is directly reconstructed by designing a linear transformation to obtain the first dataset. The type B operator is designed for clustering and linear transformation of the dataset to obtain the second new dataset. The third operator, namely, the type C operator, reconstructs the dataset by clustering and convolution to obtain the third dataset. Finally, the base classifier is trained based on the three new datasets, and then the classification results are fused by decision weighting. In the experimental section, two representative PD speech datasets are used for verification. The results show that the proposed algorithm is effective. Compared with other algorithms, the proposed algorithm achieves apparent improvements in terms of classification accuracy.
Introduction
Diagnosing Parkinson’s Disease (PD) through speech analysis is a non-invasive and promising clinical tool. However, a significant hurdle remains: intra-subject variability. Within a single patient’s recording session, some speech segments clearly exhibit symptoms like tremors or metrical issues, while others may appear healthy.
Most existing research focuses on feature engineering or classifier optimization, often treating all samples from a subject as equally valid or attempting to prune "bad" samples. This paper argues that no sample is entirely useless, nor is any single sample a perfect representation of the disease state. Simply deleting data loses information, while keeping noisy data confuses models. We propose a "sample transformation" framework that converts inconsistent raw samples into a stable, high-quality "subject envelope" representation.
Methodology: The MRCST Algorithm
The proposed Multitype Reconstruction and Clustering Transformation (MRCST) algorithm moves away from raw sample classification toward reconstructed datasets using three distinct operators:
- Operator A (Direct Linear Reconstruction): This operator generates a global statistical profile of a subject's speech by calculating six anchors: mean, median, trimmed mean (25%), standard deviation, interquartile distance, and mean absolute error.
- Operator B (Clustering Reconstruction): To capture local patterns that a global average might miss, this operator uses iterative K-means clustering to group similar segments before applying linear reconstruction to each cluster.
- Operator C (Convolution Reconstruction): This step utilizes a two-dimensional convolution to merge the clustered segments from Operator B with the linear features from Operator A. This effectively synthesizes a more robust representation by blurring the boundaries between different speech clusters.
Final classification is achieved by training base classifiers (e.g., SVM, Random Forest) on these three reconstructed datasets and fusing their predictions through a weighted decision process.
Experimental Results
The study utilized two benchmark PD speech datasets: the Sakar Dataset (40 subjects) and the MaxLittle Dataset (31 subjects). Performance was evaluated using Leave-One-Subject-Out (LOSO) cross-validation to ensure clinical generalizability.
As shown in the experimental results, the integration of MRCST consistently improved performance across different dimensionality reduction techniques (LPP, PCA, LDA) and base classifiers. On the Sakar dataset, the combination of MRCST with LDA and Random Forest achieved an accuracy of 91.25%. On the MaxLittle dataset, the same pipeline reached 94.83% accuracy, while an SVM-based approach achieved a peak of 95.17%.
When compared to existing literature, the MRCST algorithm demonstrated superior performance. For the Sakar dataset, the proposed method (91.25%) significantly outperformed previous benchmarks, which ranged from 55% to 87.5% accuracy. Ablation studies further confirmed that the fusion of all three reconstruction operators (A, B, and C) consistently outperformed any single operator, validating the multi-type approach.
Conclusion
The MRCST algorithm addresses the critical issue of sample inconsistency in PD speech datasets. By shifting the focus from sample selection to sample transformation, the "subject envelope" concept provides a more stable and accurate basis for diagnosis. Future work will explore integrating these reconstruction operators into deep learning architectures and testing the framework on larger, real-world clinical datasets.
Summary updated