Wearable devices and machine learning revolutionize Parkinson's disease monitoring

In a recent longitudinal study published in npj Parkinson's Disease, researchers tracked the quantitative progression of motor symptoms of Parkinson's disease (PD) over time using wearable sensor data and machine learning (ML) algorithms.

Study: Identification of motor progression in Parkinson’s disease using wearable sensors and machine learning. Image Credit: metamorworks/Shutterstock.com

Background

The current gold standard scale to monitor PD progression, especially motor and non-motor symptoms, is the Movement Disorder Society-Unified Parkinson's Disease Rating Scale (MDS-UPDRS).

However, variability in its assessments often hinders statistical analyses in clinical studies. Thus, a continuous interval scale is highly desirable for measuring the effectiveness of clinical interventions for PD in clinical trials.

Wearables are invaluable tools for monitoring motor symptom(s) progression in PD. They are portable, affordable, and can assess features of walking and balance spatiotemporally.

Moreover, these devices can generate in-depth and personalized kinematic measurements remotely, e.g., at homes and clinics. However, not all numerical measures extracted by wearable devices are relevant in clinical practice. Hence, ML models come into the picture.

A recent work demonstrated that the analysis of IMU data can distinguish PD patients with different severity levels and other PD-like disorders, e.g., progressive supranuclear palsy (PSP). Well-trained ML models can also identify signs of bradykinesia in PD patients. 

About the study

In the present study, researchers utilized simple Linear Regression (LR) and Random Forest (RF) algorithms with different routines of automatic feature selection to develop seven ML models and handle the wearables-measured kinematic features. 

In addition, they used walking (two minutes) and postural sway (30 seconds) data collected by six wearable inertial measurement units (IMUs) to identify the preliminary signal(s) of motor symptom progression in 74 PD patients over 18 months. In the study duration, all participants completed a total of seven visits.

Eligibility criteria mandated that these participants had PD or received anti-PD medication but did not have major musculoskeletal problems or dementia at enrollment and giving consent.

The team asked them to wear wearable sensors on their wrists, feet, sternum, and lumbar region. These devices collected triaxial accelerometer, gyroscope, and magnetometer data at 128 Hz sampling frequency.

The researchers validated the association of the wearable sensor-derived IMU data with the MDS-UPDRS-III ratings to understand which better tracked the progression of motor symptoms of PD.

They hypothesized that these models could detect a statistically significant progression of motor symptoms in PD patients earlier than the MDS-UPDRS-III scale.

Results

The researchers collected IMU data of over 18 months from 91 people with idiopathic PD. Of 122 measured kinematic features, 29 markedly linearly surged or declined at a group level over time.

Of these, 19 reflected step-to-step walking variability, previously shown to scale with disease severity in PD. Studies have also shown that it is a key predictor of falls in PD patients.

The mediolateral sway velocity was the only postural sway feature that progressed significantly; it is also a well-recognized biomarker of falls in PD patients. Among individual features, the angle of the foot at foot strike and toe-off and the stride length contributed most to the estimate of the MDS-UPDRS-III score.

A multivariate LR model (model 1) used the two kinematic features, showing the most statistically significant temporal progression. From 29 progressing features, forward feature selection identified six for use in the early stopping model (model 2). The team also investigated the RF Regressor with 29 progressing features as input (Model 3).

Applying principal component analysis (PCA) to the 122 features and 29 progressing features reduced the dimensionality of the original high-dimensionality datasets, and it returned 31 and 10 features, respectively.

Both principal components served as independent variables in LR and RF regression. It fetched models 4, 5, 6 & 7, which used LR on ten factors, RF on ten factors, LR on 31 factors, and RF on 31 factors, respectively.

The RF regressor (Model 3) estimated the MDS-UPDRS-III score with the lowest Root Mean Square Error (RMSE) (=10.02) across the five cross-validation iterations; thus, it was adopted to process the longitudinal sensor data from sequential visits. 

Model 3 also identified motor symptom progression in PD as early as 15 months after baseline, while the MDS-UPDRS scale did not capture these signs even by the end of the study period.

Furthermore, the model output increased monotonically from one visit to the next. On the contrary, the MDS-UPDRS-III scores fluctuated from visit to visit, fetching blurred evidence of the progression of PD's motor symptoms.

Conclusions

Overall, the wearables- and ML algorithms-based methodology presented in this study could be a complementary tool in clinical practice to determine early signs of PD motor symptom progression. 

This method performed better than the conventionally used clinical rating scales in PD; thus, it could dramatically improve PD patients' diagnostic and prognostic accuracy.

Written by

Neha Mathur

Neha is a digital marketing professional based in Gurugram, India. She has a Master’s degree from the University of Rajasthan with a specialization in Biotechnology in 2008. She has experience in pre-clinical research as part of her research project in The Department of Toxicology at the prestigious Central Drug Research Institute (CDRI), Lucknow, India. She also holds a certification in C++ programming.