Transformers for 1D signals in Parkinson’s disease detection and severity prediction from gait

Prerequisites

• Python 3.7
• CPU or NVIDIA GPU + CUDA CuDNN
• Install the prerequisite's libraries with pip install -r requirements.txt

Dataset

The dataset used in the paper is from Physionet. It can be downloaded from (1). A sample of the dataset is available in the subfolder data.

Getting Started

The entry point is train.py file. It has 3 arguments:

• input_data: Input folder containing the dataset.
• exp_name: type of experiment to run:
  • 'train_classifier': Run the cross-validation experiment for Parkinson detection.
  • 'train_severity': Run the cross-validation experiment for Parkinson severity prediction using the UPDRS scale.

Once executed, the algorithm will generate the following output files:

├── output (dir)
    ├── exp_name_Day_Month (when the program is launched) 
        ├── hour_minutes (when the program is launched)  
            ├──  weights.hdf5 : weights of the model
            ├──  res_seg.csv: results of accuracy, sensitivity and specificity by segments.
            ├──  res_pat.csv: results of accuracy, sensitivity and specificity by patients. 
            ├──  training_i.csv: training loss and accuracy for each fold i. For 10-fold cross-validation, 10 files. 
            ├──  model.json: Model architecture (keras)
            ├──  gt.csv: and pred.csv  ( for severity prediction) ground truth level and prediction level for each patient.
            ├──  confusion_matrix.csv: Confusion matrix for severity prediction

References

(1) https://physionet.org/content/gaitpdb/1.0.0