Abstract
Assessing the progression of movement disorders such as Parkinson's Disease (PD) is key in adjusting therapeutic interventions. However, current methods are still based on subjective factors such as visual observation, resulting in significant inter-rater variability on clinical scales such as the Unified Parkinson's Disease Rating Scale (UPDRS). Recent studies show the potential of sensor-based methods to address this limitation. The goal of this systematic review is to provide an up-to-date analysis of contactless sensor-based methods to estimate hand dexterity UPDRS scores in PD patients. Two hundred and twenty-four abstracts were screened and nine articles selected for analysis. Evidence obtained in a cumulative cohort of $n = 187$ patients and 1,385 samples indicates that contactless sensors, particularly the Leap Motion Controller (LMC), can be used to assess UPDRS hand motor tasks 3.4, 3.5, 3.6, 3.15, and 3.17, although accuracy varies. Early evidence shows that sensor-based methods have clinical potential and might, after refinement, complement or serve as a support to subjective assessment procedures.
Introduction
Clinicians currently track the progression of Parkinson’s Disease using the UPDRS, which remains the clinical gold standard. However, this scale relies heavily on visual observation, leading to significant "inter-rater variability"—where different neurologists may assign different scores to the same patient. Furthermore, these clinical assessments are episodic, capturing only a brief snapshot of a patient's condition during hospital visits. There is a critical need for objective, reproducible, and potentially remote methods to quantify hand dexterity, specifically tremors and bradykinesia (slowness of movement), to better adjust therapeutic interventions like dopaminergic replacement therapy.
Methodological Framework
This systematic review followed PRISMA guidelines to identify high-quality studies utilizing contactless sensors for motor assessment. The primary technology identified was the Leap Motion Controller (LMC), a small peripheral that uses infrared light to track hand and finger points in 3D space.
The methodological pipeline across the reviewed studies generally followed a standardized pattern for translating raw sensor data into clinical scores:
- Data Capture: Recording patients performing specific MDS-UPDRS III tasks (e.g., finger tapping).
- Feature Extraction: Calculating spatial variables such as Euclidean distance between fingertips or palm center coordinates.
- Signal Processing: Applying techniques like low-pass filtering to isolate Parkinsonian tremors (typically 4–6 Hz) from background noise.
- Modeling: Utilizing machine learning for either classification (assigning discrete UPDRS scores) or linear regression (predicting continuous severity).
Key Insights and Results
The review synthesized data from a cumulative cohort of 187 patients and 1,385 individual task samples. Several critical insights emerged regarding the efficacy of sensor-based quantification:
- The "Aha" Variables: The most reliable predictors of bradykinesia are the opening and closing speeds of the fingers. These features showed the strongest correlations with clinical scores ($R > 0.8$ in some studies), proving more informative than simple repetition counts.
- Tremor Localization: For resting tremors, tracking palm center coordinates proved to be a more stable and accurate predictor than tracking individual fingertips.
- Task Sensitivity: While finger tapping (task 3.4) is highly quantifiable, Pronation-Supination (task 3.6) remains difficult for current contactless sensors to track accurately due to hand self-occlusion during wrist rotation.
In terms of performance, tremor classification achieved up to 97% accuracy in specific models. For bradykinesia, classification accuracies for finger tapping ranged between 82% and 84% using Support Vector Machines (SVM).
Discussion and Future Directions
While the evidence suggests that LMC-based systems are nearly as accurate at classifying PD symptoms as they are at distinguishing patients from healthy controls, several hurdles remain. Most current studies lack data on patients with severe symptoms (UPDRS scores 3 or 4), and infrared sensors remain sensitive to ambient light conditions.
A major validation challenge is that the "ground truth" (human rating) is itself imperfect. Future research must determine if the sensor's classification error falls within the range of inherent inter-rater variability among clinicians. Moving forward, these objective variables—such as power spectral densities and movement speeds—could set the basis for "expert systems" that provide a finer resolution of disease progression than the current 0–4 integer scale allows.
Summary updated