Machine Learning Models for Parkinson's Disease Gait Assessment and Medication Adherence from Smartphone Sensor Data

Abujrida, Hamza

Etd

Machine Learning Models for Parkinson's Disease Gait Assessment and Medication Adherence from Smartphone Sensor Data

Public Deposited

Parkinson’s Disease (PD) is a neurodegenerative chronic disorder with multiple motor and non-motor symptoms. People afflicted with Parkinson's Disease experience severe problems with performing daily activities including their gait (the way a person walks), which frequently links to a less steady walk, that arises from changes in posture, slowness of movement (bradykinesia), and a shortened stride. This distinctive walk is called ‘Parkinsonian gait. When a PD patient develops a Parkinsonian gait, they start to experience festination: progressively shorter but accelerated steps forward, often in a shuffling manner. Other symptoms include slowness of gait, hesitation of starting gait aka Freeze of Gait (FoG), difficulty making turns, and postural instability leading to frequent falls. Some features of Parkinsonian gait are likely to become more pronounced over time, particularly festination, stooped posture, and FoG. PD patients feel unsteady and lose confidence because of the fear of falling. Consequently, their social activities and their quality of life get severely impacted. As PD has no ultimate cure, physicians aim to delay PD complications, especially those that degrade the patient’s quality of life such as motor symptoms and dyskinesia. Patients' lack of adherence to prescribed medication is a major challenge for physicians, especially for patients suffering from chronic conditions. The Centers for Disease Control and Prevention (CDC) estimates that medication non-adherence causes 30 to 50 percent of chronic disease treatment failures and 125,000 deaths per year in the USA [119]. In PD patients particularly, adherence varies between 10% and 67% [120]. Since changes in PD gait can be a good measure for inferring the progression and severity of the disease to inform early intervention, gait has been part of the motor section of the Movement Disorder Society-Sponsored Revision of the Unified Parkinson’s Disease Rating Scale (MDS-UPDRS). MDS-UPDRS motor part is mostly assessed by a professional clinician. To reduce the cost of care, remote assessment in patients’ homes has recently become an alternative tool for monitoring the progression of Parkinson’s disease (PD). Smartphones, particularly, provide an affordable, accessible, and easy-to-use platform for PD gait sensing. Smartphones are ubiquitous, portable, and user-friendly. Equipped with triaxial Accelerometers and gyroscopes in addition to powerful CPUs, smartphones offer the potential for remote gait assessment in the patient’s home environment. In this dissertation, to facilitate an accurate remote Parkinsonian gait assessment, we proposed and rigorously evaluated a novel Deep-Learning (DL) based gait analysis system that assesses the severity of PD gait based on 30-seconds walk given by the patient, facilitating home-based clinical monitoring by remotely assessing the PD patient gait. Specific preprocessing steps were utilized including the calculation of the moving average, subtracting signal mean, and detection of gait strides. These techniques resulted in smoothing the signal, filtering of noise, and cancellation of gravity/breeding effects. These steps facilitate DL automatic feature extraction and eliminated the need for any kind of signal conversion. The most significant contribution of our work is the proposal of a deep-learning-based system that comprehensively classifies 3 PD symptoms: the severity of FoG, walking imbalance, and shaking/tremors from data gathered in one study. Prior work has trained and tested separate models to analyze each of these PD gait anomalies separately, the model we introduced is a single model that achieved impressive results for all of the PD gait symptoms. This was challenging because the model’s parameters had to be jointly tuned in order to establish relationships with different sets of PD symptom labels, all while using the same dataset as an input. To achieve the ultimate results, we investigated four different approaches based on multiple Machine Learning (ML) algorithms. The first approach employs the extraction of hand-crafted features as input to Machine learning algorithms, we conducted supervised classification experiments using 10-fold cross-validation and measured the performance of different models. In the second approach, we encoded the walking signal to an image format using the Gramian Angular Field (GAF) encoder. We employed the concept of Transfer Learning on the top image-based models such as ResNet50, Inception, SqueezeNet, and EfficientNet. The third approach employed variations of the Long Short Term Memory models, we investigated the simple LSTM, CNN-LSTM, and parallel LSTM models. These first three models had limitations, necessitating research and development of a fourth method. The first ML approach, could not achieve an acceptable performance when classifying various walks, mainly because the handcrafted features were not able to linearly or non-linearly discriminate between the different classes. The second and Third approaches suffered data overfitting, because of the models' over-complexity, which could not be justified by our dataset, due to a large number of trainable parameters. Consequently, a DL multi-layer Conventional Neural Network (CNN) model was introduced, this model operates on 1Dimensional convolution filters to classify 30 seconds of walking data into one of five severity levels. Our DL network was able to Classify the PD Walking-Balance, Shaking/Tremor, and Freeze of Gait (FoG) symptoms, with an accuracy of: 99.1%,98.4%, and 98.2% respectively. Another important contribution of this work is the model's ability to discriminate between PD patients on- vs off-medication and baseline HC walk. Unlike methods such as Drug-Bottles and urine or blood test that monitor discrete medication-related events, our approach analyzes data corresponding to continuous windows of time, submitted by PD patients every time they walk before/after taking their medication. By training our model on walking segments recorded before and after medication, we were able to present a medication adherence system that operates with an accuracy of 98.2%, which facilitates remote medication adherence. Finally, our DL-based gait analysis system was successfully applied to more than 450 participants from the independent dataset (mPower dataset). This system is proven to be applicable in home environments and capable of providing an accurate PD gait assessment in a telemedicine fashion.

Creator