Quantifying Severity of Symptoms in Patients with Parkinson’s Disease Using Wearable Sensors and Machine Learning

Abstract

Parkinson’s disease has been an active area of research from a long time. Although no cure for it has been discovered yet, but certain medicines, surgical treatments, or other therapies can sometimes provide some relieve with some of the symptoms. Most prominent Parkinson’s symptoms include Tremors, Dyskinesia, and Bradykinesia. Detection and prediction of these symptoms holds big significance as it can help in avoiding accidents, injuries etc., by generating alerts or generating cues to help reduce amplitude of anomaly. Our research focuses on detection of Parkinson’s and on quantifying severity of Parkinson’s symptoms using inertial data. For this purpose, we have used inertial dataset shared in the levodopa study, where inertial data is collected from three different body locations i.e., upper two limbs and waist, using 3 different sensors i.e., An IMU (GeneActiv), a smart watch (Pebble) and a smartphone (Samsung Galaxy S2). We present a 2-stage anomaly detection and classification pipeline where in the first stage (binary classification) an input signal segment is classified as either anomalous or as a normal signal. If the input signal is anomalous, then it passes through second stage (multi-label classification) where we categorize the type of anomalies present in the signal and thus quantify the severity of symptoms. We also present a performance comparison by using different preprocessing hyperparameters (segment sizes and segment labelling methodologies), of machine learning (Random Forest) vs deep learning (HARDenseRNN) based techniques. Overall, we present results on 3 different datasets (sourced from 3 different sensor [GeneActiv IMU, Pebble smartwatch, Samsung Smartphone]), where for each dataset we make 6 different preprocessing configurations by using 3 different segment sizes (50, 150, and 250 datapoints) and 2 different segment labeling methodologies (using mode as the segment label and identifying signal as anomalous if it contains any anomalous portion in it). Then we have 2 models for machine learning and 2 for deep learning, one for each of the two stages of pipeline, for each of which we present a detailed comparative performance analysis, for each of the six-preprocessing configuration and for all three sensors. For machine learning models we are using feature engineering on raw data and present top performing features as well, whereas we feed raw inertial data (magnitude of raw values) to the deep learning models.