Hybrid CNN-BiLSTM with Attention Mechanism for Interpretable Parkinson’s Disease Detection from Speech Signals
Research Digest on Engineering Management and Social Innovations · 08.04.2026 · 2 (4)
DOI: 10.46647/icetetas156 ↗Kurz gefasst
Fachartikel in der Zeitschrift Research Digest on Engineering Management and Social Innovations, online erschienen am 08.04.2026. Verfasst von Sakinala Gopal und 4 weiteren Autor:innen.
Abstract (Originalsprache)
Background
Parkinson’s disease (PD) affects ~10 million people globally; 50–60% of dopaminergic neurons are lost before motor diagnosis. Acoustic speech disturbances emerge 6–24 months before motor symptoms in up to 90% of patients, offering a non-invasive early biomarker.
Methods
PFT-HCLA, a Partially Fine-Tuned Hybrid CNN-BiLSTM-Attention architecture, extracts a 55-dimensional feature vector (40 MFCC + 12 Chroma + Spectral Centroid + Rolloff + ZCR) from 831 recordings (431 PD, 400 HC) of the Italian Parkinson’s Voice and Speech Database. Dual Conv1D blocks provide spectral extraction; Bidirectional LSTM models temporal dynamics; scaled dot-product attention yields frame-level interpretability. Layers 1–6 are frozen; layers 7–12 are fine-tuned to prevent catastrophic forgetting.
Results
95.1% accuracy (95% CI: 93.2–96.8%), 95.6% sensitivity, 94.0% specificity, F1 = 0.954, AUC-ROC = 0.962. Early-stage (HY-1) detection: 90.3%. Attention contributes +3.5 pp over CNN-BiLSTM baseline (p < 0.001).
Conclusion
PFT-HCLA sets a new benchmark for speech-based PD detection with clinically interpretable attention aligned to established dysarthria biomarkers.
Zitieren
APA
Gopal, S., Hema, K., Roshini, A., Lakshmi, S. S., & Suvarchala, P. N. V. (2026). Hybrid CNN-BiLSTM with Attention Mechanism for Interpretable Parkinson’s Disease Detection from Speech Signals. Research Digest on Engineering Management and Social Innovations, 2(4). https://doi.org/10.46647/icetetas156
Vancouver
Gopal S, Hema K, Roshini A, Lakshmi SS, Suvarchala PNV. Hybrid CNN-BiLSTM with Attention Mechanism for Interpretable Parkinson’s Disease Detection from Speech Signals. Research Digest on Engineering Management and Social Innovations. 2026;2(4). doi:10.46647/icetetas156
BibTeX
@article{Gopal2026,
author = {Gopal, Sakinala and Hema, Koppanathi and Roshini, Akula and Lakshmi, Seelam Sri and Suvarchala, Pasupuleti Naga Venkata},
title = {Hybrid CNN-BiLSTM with Attention Mechanism for Interpretable Parkinson’s Disease Detection from Speech Signals},
journal = {Research Digest on Engineering Management and Social Innovations},
year = {2026},
volume = {2},
number = {4},
doi = {10.46647/icetetas156}
}
EUVIMED