This project utilizes a dataset from Kaggle containing 1500 Parkinsonian voice recordings, and for each voice recording, a labelled symptom severity score (UPDRS) is provided. A variety of regression models are fitted on the dataset to predict the UPDRS score. Hyperameter tuning is applied to each model and model performance metrics are compared. See the medium article below to learn more about the project and the model building steps taken.
Medium article link: https://medium.com/@bobbywilt1/predicting-parkinsonian-symptom-severity-from-voice-recordings-5efcfdc406a
Dataset obtained from: https://www.kaggle.com/mountainguest/parkinsons-telemonitoring?select=parkinsons_updrs.names
Original research paper: https://www.researchgate.net/publication/40026354_Accurate_Telemonitoring_of_Parkinson%27s_Disease_Progression_by_Noninvasive_Speech_Tests
Assessing Parkinson's symptom severity is typically completed during in-person clinical visits. Due to the Covid pandemic and travel difficulties for patients, attending an in-person visit can be problematic for patients. Remote-accessible diagnosistic methods would be preferrable for some Parkinson's patients and may even help patients and clinicians track a patient's symptoms more closely than in-person assessments. A research study by Dr. Tsanas was conducted in 2009 to assess the accuracy of using voice recordings to predict a patient's disease symptom severity score (UPDRS). The goal of this project is to develop a more accurate machine learning model than the model mentioned in the research paper above by using the same dataset and by implementing hyperparamter tuning.
Download and run the notebook file "PD_Process" using your text editor of choice.
Bobby Wilt - Linkedin