Diagnosing Parkinson’s

1. by: Peter Hirschmann Diagnosing Parkinson’s

2. Diagnosing Methods Traditional Testing • Monitor symptoms such as: • Resting Tremor • Bradykinesia • Rigidity • Postural Instability • Sub-symptom • Voice Problems • Use classification teaching algorithms to identify Parkinson’s

3. Parkinson’s Disease • “Movement disorder that is chronic and progressive” Parkinson's Disease Foundation • There is currently no cure • Treatment involves surgery or medication Parkinson Disease

4. Data – UCI Machine Learning Repository Label Vectors • MDVP:Fo(Hz) - Average vocal fundamental frequency • MDVP:Fhi(Hz) - Maximum vocal fundamental frequency • MDVP:Flo(Hz) - Minimum vocal fundamental frequency • MDVP:Jitter(%),MDVP:Jitter(Abs),MDVP:RAP,MDVP:PPQ,Jitter:DDP – Several measures of variation in fundamental frequency • MDVP:Shimmer,MDVP:Shimmer(dB),Shimmer:APQ3,Shimmer:APQ5,MDVP: APQ,Shimmer:DDA - Several measures of variation in amplitude • NHR,HNR - Two measures of ratio of noise to tonal components in the voice • RPDE,D2 - Two nonlinear dynamical complexity measures • DFA - Signal fractal scaling exponent spread1,spread2,PPE - Three nonlinear measures of fundamental frequency variation • Status - Health status of the subject (one) - Parkinson's, (zero) - healthy Feature Vectors LOO • Leave One Out – Useful for realistic testing, since all known data would be used for testing new patients.

5. Classification Methods • Polynomial Mode • Tries to fit the known data to a polynomial model • Maximum Likelihood • Provided the mean and variance, it bases the identified vectors on a chosen model, Gaussian, and attempts to fit model parameters • Nearest Neighbor • Searches and labels vectors based on the closest identified vectors

6. Results – Polynomial Model • Training Error - Blue • Testing Error - Green • LOO Error - Red • Clearly, LOO has the lowest Sum of Square Error Sum Square Error Features 1-22

7. Results – Maximum Likelihood • Training Data increases with x-axis and Testing Data Decreases • LOO testing method Classification Rate Samples: 1-185 Classification Rate Samples: 1-195

8. Results – Nearest Neighbor • 7 neighbors, Classification Rate vs. Percentage of Data as Testing Data • LOO Method, Classification Rate vs. # of Neighbors Classification Rate Percentage of Data used as Testing Data 1%-95% Classification Rate # of Neighbors 1-7

9. Summary • Polynomial Model • This classification method only proves that single feature vectors are not adequate but that LOO is best way to train • Maximum Likelihood • Max Percentage when Training vs. Testing Data = 0.8203 • Classification Rate using LOO = 0.9330 • Nearest Neighbor Neighbors: 1 2 3 4 5 6 7 C_Rate: 1.0 0.8814 0.8763 0.8711 0.8711 0.8454 0.8247 • Best Classification Rate occurs when over 95% of the data is for training for every amount of neighbors • When using LOO, the best results occur from one nearest neighbor

10. Conclusion • Nearest Neighbor using LOO methodology with 1-7 neighbors are all above 80% • KNN is the best method of the three tested • Still does not beat the comfort and knowledge of a doctor, but highly useful as a tool for mass testing. Would not require a M.D. to eliminate people the majority of people who do not have Parkinson’s