MUMBAI, India, Oct. 26 -- Intellectual Property India has published a patent application (202641123904 A) filed by Srm Institute Of Science And Technology, Ramapuram Campus; and Easwari Engineering College on October 16, 2026, for System And Method For Early Detection Of Parkinson’s Disease Using Multiscale Attention-Based Speech Analysis.
Inventors include Dr. J. Sutha; Y. Vighnesh; Arya Simlot; and Neil Nayak.
The application for the patent was published on October 23, 2026, under issue no. 43/2026.
Abstract: Abstract: Despite the fact that the essence of the research consists in the early detection of Parkinson through the invention of voice, it is a mixture of the raw voice measurements and the extra-detailed time-specific patterns of Mel-spectrograms. The method, which is based on a Time-CNN model, processes these spectrograms with the use of spread-out 2D convolutions initially followed by the transition to sequential 1D processing. Audio samples were collected between 131 people in early stages of PD and 42 participants without the disorder - these included holding of the vowel sounds and syllable repetition. The team reduced them using principal component analysis, rather than by using only the conventional measures as MFCC, jitter, shimmer, spectral centroid and pitch (F0). Dimensionality reduction also simplified the structure of inputs prior to classification instead of handling them separately. Hybrid feature fusion has been shown to be more effective than single- feature tests, with an accuracy of 0.78 and F1-score of 0.831. When using spectral data together with time-based deep learning features, the results are more consistent and do not vary as much even with a change of the structure of the neural network. Still, there are some problems. One of the problems: the tapeings are based in closed, noisy rooms and this decreases the usefulness of treatment out of the laboratory. The other problem is that the design does not literally depict several frequency scales, and it may be missing some subtle spectral irregularities. Instead of dynamically adjusting the weights, the mixture of raw and learned features is based on simple stacking. Instead of trying to find uniform patterns in devices, the procedure employs mean values as a short-cut. This is despite the visual representation, such as spectrograms, which make it somewhat limited to understand what is being decided inside. Tests on performance on various languages or spontaneous inputs on speech have not been done yet. This is a new step: Multiscale Attention-Fusion TemporalNet Framework is the first attempt to fill existing gaps by introducing a blended model based on Time-CNN but having several convolution streams at different levels. Instead of outputpiling, fusion occurs through attention-based weighting - prioritising where it is most needed. Low tones and sharp variations of higher ranges are caught at a multi-level spectrogram analysis module. Since microphones vary in nature, the layers created to adapt to different types of microphones assist to stabilize performance irrespective of the recording equipment used. This organization is expected to achieve closer divisions of classes, more consistent Matthews correlation outcomes on validation splits, and enhanced practicality in the real world, in terms of detecting early-stage Parkinsonian voice characteristics.
Disclaimer: Curated by HT Syndication.
