In a recent longitudinal study published in npj parkinson’s diseaseresearchers used wearable sensor data and machine learning (ML) algorithms to track the quantitative progression of motor symptoms in Parkinson’s disease (PD) over time.
Research: Identifying motor progression in Parkinson’s disease using wearable sensors and machine learning. Image credit: metamorworks/Shutterstock.com
background
The current gold standard scale for monitoring PD progression, particularly motor and non-motor symptoms, is the Movement Disorder Society Unified Parkinson’s Disease Rating Scale (MDS-UPDRS).
However, variability in assessment often hinders statistical analysis in clinical studies. Therefore, continuous interval scales are highly desirable for measuring the effectiveness of clinical interventions for PD in clinical trials.
Wearables are an invaluable tool for monitoring the progression of motor symptoms in PD. They are portable, affordable, and allow for the assessment of gait characteristics and spatiotemporal balance.
Furthermore, these devices can generate detailed and personalized kinematic measurements remotely, such as at home or in the clinic. However, not all numerical indicators extracted by wearable devices are relevant to clinical practice. That’s where ML models come in.
Recent studies have demonstrated that analysis of IMU data can differentiate between PD patients with different severity levels and other PD-like disorders such as progressive supranuclear palsy (PSP). A well-trained ML model can also identify signs of bradykinesia in PD patients.
About research
In this study, researchers utilized simple linear regression (LR) and random forest (RF) algorithms, along with various automatic feature selection routines, to develop seven ML models to improve the kinematics measured on wearables. Processed features.
Additionally, we used gait (2 min) and postural sway (30 s) data collected by six wearable inertial measurement units (IMUs) to investigate motor symptom progression in 74 PD patients over 18 months. I have identified a preliminary signal. All participants completed a total of seven visits during the study period.
Eligibility criteria required that these participants had PD or were receiving anti-PD medications, but did not have significant musculoskeletal problems or dementia at the time of enrollment and consent. .
The researchers asked participants to wear wearable sensors on their wrists, feet, sternum, and lower back. These devices collected 3-axis accelerometer, gyroscope, and magnetometer data at a sampling frequency of 128 Hz.
Researchers examined the association between wearable sensor-derived IMU data and MDS-UPDRS-III assessments to understand which one better tracks the progression of motor symptoms in PD.
They hypothesized that these models may be able to detect statistically significant progression of motor symptoms in PD patients earlier than the MDS-UPDRS-III scale.
result
Researchers collected over 18 months of IMU data from 91 patients with idiopathic PD. Of the 122 kinematic features measured, 29 significantly increased or decreased linearly at the group level over time.
Of these, 19 reflected stage-to-stage gait variation, which has previously been shown to be proportional to disease severity in PD. Studies have also shown that this is an important predictor of falls in patients with PD.
Medial and lateral sway velocity was the only postural sway characteristic that significantly improved. It is also a well-known biomarker of falls in PD patients. Among the individual characteristics, foot angle during landing and toe-off, and stride length contributed the most to the estimation of his MDS-UPDRS-III score.
The multivariate LR model (Model 1) used two kinematic features and showed the most statistically significant temporal progression. From the 29 ongoing features, forward feature selection identified 6 features to use in the early stopping model (Model 2). The team also investigated an RF regressor (Model 3) that used 29 progressive features as input.
Applying principal component analysis (PCA) to 122 features and 29 in-progress features reduced the dimensionality of the original high-dimensional dataset and returned 31 and 10 features, respectively.
Both principal components served as independent variables in the LR and RF regressions. I fetched models 4, 5, 6, and 7. Models 4, 5, 6, and 7 used LR with 10 factors, RF with 10 factors, LR with 31 factors, and RF with 31 factors, respectively.
The RF regressor (Model 3) estimated MDS-UPDRS-III scores with the lowest root mean square error (RMSE) (=10.02) across five cross-validation iterations. Therefore, it was employed to process longitudinal sensor data from consecutive visits.
Model 3 also identified progression of PD motor symptoms as early as 15 months from baseline, whereas the MDS-UPDRS scale did not capture these signs by the end of the study period.
Additionally, the model output increased monotonically with each visit. On the contrary, MDS-UPDRS-III scores varied from visit to visit, providing equivocal evidence regarding the progression of motor symptoms in PD.
conclusion
Overall, the wearable and ML algorithm-based methodology presented in this study may serve as a complementary tool to determine early signs of PD motor symptom progression in clinical practice.
This method showed better performance than clinical rating scales traditionally used in PD. Therefore, it has the potential to dramatically improve the diagnostic and prognostic accuracy of PD patients.