Analyzing movement patterns to facilitate the titration of medications in late stage Parkinson's disease

Parkinson's disease is the second most common neurodegenerative disorder after Alzheimer. Nowadays more than 10 million people worldwide are affected and this number is increasing at a rate of 60 thousands new diagnoses per year in US only. It has an impact on the US healthcare industry of 23 billion dollars per year represented by direct and indirect costs. Periodically quantify the severity of the symptoms is important to arrange the medication doses and schedule the intake times to avoid or keep at minimum the side effects of the medications and keeping always a low level of impairment caused by the symptoms. This is a problem of drug titration, typical of those drugs that have a wearing off effect, like Levodopa, used for Parkinson's Disease. Today to quantify the severity of motor symptoms, simulated tasks are performed by patients under the observation of clinicians who patiently assigns scores to each task. These procedures take a lot of time to both patients and clinicians and represent an important part of the total cost upon the healthcare industry. Wearable sensors and modern techniques of data analysis and Artificial Intelligence can help in this task by predicting the severity of motor and non motor symptoms, allowing for continuous monitoring, objective analysis and saving time and money to patients and hospitals. The Motion Analysis Laboratory (Harvard Medical School) collected data from 27 patients with Parkinson's Disease, during performed tasks in the laboratory environment and simulated activity of daily life (SADL) in an apartment-like environment. This thesis is part of a bigger project where different hospitals and universities in Boston, MA are involved called BlueSky project. Aim of this project is to use all the information from all different kind of sensors (accelerometric and physiological data) to accurately predict changes in the severity of Parkinson for both motor and non-motor symptoms. The main focus of the thesis is on the analysis of accelerometric data acquired from wrists and feet. However the first months have been spent analyzing other physiological data such as Galvanic Skin Response and Heart Rate Variability for the study of non motor symptoms. Finally, a machine learning approach has been used. Different regressor models have been tested, the focus has been on the study of temporal patterns through Recurrent Neural Networks, the overall error is reported in terms of Root Mean Squared Error as used in regression problems, however the characteristics of the classifier are derived from the regressor to a better comparison among different models. The best results obtained are using a LSTM (Long-Short Term Memory) network on the output of the RF (Random Forest) leading to an overall accuracy of 81.4% (4 classes) and RMSE of 0.55. However this approach is better only in the laboratory dataset, while in the apartment dataset the only Random Forest has better performances.