Symposium

Of Rising ScholarsFall 2022

Pavithra will be presenting at The Symposium of Rising Scholars on Saturday, September 24th! To attend the event and see Pavithra's presentation,

Register here!
Go to Polygence Scholars page
Pavithra Kamatchi Soundaram's cover illustration
Polygence Scholar2021
Pavithra Kamatchi Soundaram's profile

Pavithra Kamatchi Soundaram

Dublin High SchoolClass of 2023Dublin, California

About

Projects

  • "Audio Signal Processing Tool for Machine Learning Enabled Detection of Parkinson’s Disease at Early Stages" with mentor Sejal (May 23, 2022)

Project Portfolio

Audio Signal Processing Tool for Machine Learning Enabled Detection of Parkinson’s Disease at Early Stages

Started Dec. 16, 2021

Abstract or project description

Parkinson’s disease (PD) is a highly prevalent neurodegenerative disorder in which neurons in the basal ganglia and substantia nigra die. The complications of PD are rated as the 14th cause of death in the U.S. in which speech deterioration and impaired movement are one of the first symptoms. With no single test to definitively diagnose PD, it is one of the most difficult disorders to identify. Many times, PD patients have only been diagnosed after 2 years of having symptoms, and 30% of the time, PD has been misdiagnosed. Since PD worsens as you age and there is no cure, it is imperative to identify the disease at a very early stage. Currently, doctors use DaTSCANS and SPECT scans which cost around $2,500 - $5,000 [18] and have a low accuracy rate of around 38% for PD detection [9].The aim of this study is to find a new machine-learning-based approach to detect PD from speech recordings at early stages without having to pay for expensive imaging tests. In order to do this, the MDVR_KCL dataset containing speech recordings of PD patients and healthy controls from Britain was used. Audio data manipulation and data wrangling by splitting the audio files into 20 second sections to increase the sample size were then used to preprocess the speech recordings. Then, using the recordings, acoustic features such as jitter and shimmer were extracted which were then used to train the machine learning model for PD detection. After using an ensemble of machine learning algorithms such as Support Vector Machines, Random Forests, Logistic Regression and the XGBoost algorithm, the XGBoost and Random Forests model was determined to have the highest classification accuracy of 93.10% and is shown to have a higher accuracy rate than the imaging tests that doctors currently use.