Devpost:https://devpost.com/software/spiro

• Python
  • Flask
  • Keras
  • Pandas
  • Numpy
  • Matplotlib
• Swift
• Google Cloud Platform
• Firebase

• Convert points into images
• Train and run machine learning algorithm
• App where a user can take the spiral test
• Return predictions to the app
• Current Model:
  • 32 Convolutions
  • 10x10 size
  • 40x40 pools
  • 0.2 dropout
  • 128 dense
  • sigmoid activation
  • 20 batch size
  • 100 epochs

• Front End
  • Plot r vs. theta graph to gain another angle on the deviation
• Back End
  • Calculate a "score" by using the predicted probability of the spiral being generated by someone with Parkinson's
  • Retrain model when enough new data is aggregated in the Firebase database

• Treat data as image with pixel values vs. time series of the spiral path points
• Pixel value from 500x500 to 125x125
• The model initially assumed that everyone had Parkinson's, so it would have a 75% accuracy due to the makeup of our initial dataset.
• For the machine learning model, we duplicated the data points of the control so we would have a 50-50 split between control and Parkinson's points. This should not contribute to overfitting because of the dropout, which adds a random perturbation in the model.
• There are many images in which the model has a better detection than human intuition, as seen in has_parkinsons.png and does_not_have_parkinsons.png