Parkinson's Disease Dyskinesia Dataset

The ALAMEDA_PD_dyskinesia_dataset.csv contains 92 features extracted from raw accelerometer data after pre-processing, 2 dyskinesia-related labels and some other metadata. In total, it includes 97 columns:

1. The first two columns correspond to the start_timestamp and the end_timestamp of the time window from which the respective features have been extracted.

2. The third column corresponds to the subject_id, which is used to uniquely identify PD patients enrolled in the current study.

3. The next 92 columns correspond to features extracted from raw triaxial accelerometer data collected with the GENEActiv smart bracelets throughout 30-min MDS-UPDRS assessment during in-clinic visits, after applying some preprocessing steps. First, the accelerometer signals were band-pass filtered [0.25 Hz, 3.5 Hz] to enable dyskinesia detection. Then, the magnitude and the first principal component of the filtered signals were computed to attenuate the dependency on sensor placement and orientation. Finally, the transformed signals were segmented into time windows of 2048 samples (or 20.48 sec) with 50% overlap. Then, 92 features were extracted in both time and frequency domains. Spectral features were extracted after applying Fast Fourier Transform. The full list of the extracted features is demonstrated in the table below. These features can feed Machine Learning models to predict the presence/absence of PD dyskinesia.

   4. The final 4 columns correspond to tremor-related labels (Constancy_of_rest, Kinetic_tremor, Postural_tremor and Rest_tremor). They derive from the respective MDS-     UPDRS III annotations, after transforming them to make them suitable for binary classification. More specifically, zero scores remained 0 to indicate the absence of tremor while positive scores were transformed to 1 to indicate the presence of tremor. Each of these columns can be used as a target to be predicted with the help of Machine Learning models.