Figure 4.

Structure of the Parkinson level classification model with mixed input data. The temporal input data (upper branch) is a moving window of 64 timestamps with the three axes of each sensor (accelerometer and gyroscope); this branch of the model is composed with a series of convolutional layers and LSTM to automatically extract the temporal characteristics of the signals. The branch with the frequency information (center branch) is the spectrogram image of the temporal signal, this branch is composed of convolutional layers to extract the information contained in the images. The branch with biomechanical variables (the lower branch) is composed of densely connected layers. All these branches are joined before the Top Model with a linear output layer between 0 and 1 with the points of 0.33 and 0.66 for the different levels.