Modeling and evaluating Parkinson's disease from movement data

Parkinson's disease is currently evaluated with the UPDRS (Unified Parkinson’s Disease Rating Scale). This master thesis was focused on four exercises from the motor examination part of the UPDRS, namely, the Finger Tapping, the Pronation Supination, the Postural Tremor, and the Kinetic Tremor tasks. This rating scale is, however, subject to the bias and the variability of the clinician who assesses the exercises. For this reason, a tool made of accelerometers and gyroscopes has been developed to record the patient's hand movement while performing the exercises. A study has, thus, been conducted where the patients performed the exercises while their hand movements were recorded and three experts rated the patients doing their exercises through video recordings. The goal of this master thesis was therefore to use the recordings of the movement data to predict the patient’s UPDRS scores and associate specific features of the movements with the UPDRS scores. To achieve this goal, the data collected were organized and preprocessed. Afterward, features were extracted from the signals. The preprocessing pipeline was configured accordingly to the steps and the parameters that were maximizing the agreement between the model output and the different annotators. Then, the performances of models trained with different subsets of features were assessed in order to determine which kind of features or which signals were the most informative for the models. Finally, the performances of the models were compared to the performances of the experts and it has been shown that the models were achieving good results overall but not as good as the experts did. However, in the case of the Kinetic Tremor, the model outperformed one expert and nearly outperformed another.