EUVIMEDEUVIMED entdecken
Crossref

Dictionary-normalized scale wise feature representation based discriminative adaptive residual sparse transformer for Parkinson’s disease detection from speech

AD Jeyarani

Biomedical Signal Processing and Control · 2026-07 · 120 · 110036

DOI: 10.1016/j.bspc.2026.110036 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Biomedical Signal Processing and Control, 2026-07. Verfasst von AD Jeyarani. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 44 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Parkinson’s disease (PD) detection is essential for enabling early intervention, effective disease management and improving patient quality of life. Speech analysis provides a promising non-invasive diagnostic path, since PD often manifests through subtle vocal impairments that can be objectively quantified. Nevertheless, reliable speech-based PD detection is challenged by variability in recording environments, speaker accents, speaking styles and channel distortions that mask disease-specific cues. In addition, high intra-class variability, limited labeled pathological data, class imbalance between healthy and PD subjects and high-dimensional acoustic features increase overfitting risk and reduce generalization. This paper presents a novel framework termed Dictionary-Normalized Scale-wise Features with Discriminative Adaptive Residual Sparse Transformer (DiNAS DART) to improve PD classification performance. First, scale-wise integration of spectral static and temporal delta and delta–delta components is performed using multi-scale Mel-Scale Frequency Cepstral Coefficients (MSFCCs) to comprehensively capture PD-related vocal characteristics. Next, an Adaptive Residual Enhanced Discriminative Sparse Transformer (AREDST) is employed combining residual learning with sparse attention to emphasize task-relevant acoustic patterns while suppressing redundant information. To further address recording and speaker variability, a dictionary-normalization strategy is incorporated to stabilize feature representations across heterogeneous conditions. Moreover, two complementary discriminative sparse coding methods namely Least Absolute Shrinkage and Selection Operator (Lasso L1) and Discriminative Non-negative Matrix Factorization (DNMF) are jointly applied to enforce sparsity, reduce feature redundancy, and enhance class separability. Experiments on the PC-GITA dataset show that DiNAS-DART achieves 99.89% accuracy, 99.78% F1-score, and 0.93 MCC with AUC values above 0.98 across all tasks demonstrating robust and superior discriminative performance.

Zitieren

APA

Jeyarani, A. (2026). Dictionary-normalized scale wise feature representation based discriminative adaptive residual sparse transformer for Parkinson’s disease detection from speech. Biomedical Signal Processing and Control, 120, 110036. https://doi.org/10.1016/j.bspc.2026.110036

Vancouver

Jeyarani A. Dictionary-normalized scale wise feature representation based discriminative adaptive residual sparse transformer for Parkinson’s disease detection from speech. Biomedical Signal Processing and Control. 2026;120:110036. doi:10.1016/j.bspc.2026.110036

BibTeX
@article{Jeyarani2026,
  author = {Jeyarani, AD},
  title = {Dictionary-normalized scale wise feature representation based discriminative adaptive residual sparse transformer for Parkinson’s disease detection from speech},
  journal = {Biomedical Signal Processing and Control},
  year = {2026},
  volume = {120},
  pages = {110036},
  doi = {10.1016/j.bspc.2026.110036}
}
Kostenlos starten →