Using an unbiased symbolic movement representation to characterize Parkinson’s disease states

Avner Abrami, Stephen Heisig, Vesper Ramos, Kevin Thomas, Bryan Ho, Vittorio CaggianoScientific 2020 · DOI

TL;DR

This paper introduces a novel method to analyze human movement by breaking it down into "syllables" and representing their sequences as a statistical grammar. This "Symbolic Movement Representation" (SMR) can objectively quantify Parkinson's disease (PD) severity, offering a more continuous and less intrusive assessment than current clinical methods, applicable both in-clinic and at-home.

What Problem Does This Paper Address?

Parkinson's disease (PD) is a neurodegenerative disorder affecting motor control, characterized by tremor, rigidity, and bradykinesia. Current diagnosis and monitoring rely heavily on the clinician-administered Movement Disorder Society Unified Parkinson's Disease Rating Scale (MDS-UPDRS). This scale has several limitations: it's administered infrequently (once or twice a year), can be influenced by patient stress (e.g., "white coat effect"), medication cycles, and inter-rater variability among clinicians. Furthermore, it requires patients to perform specific, scripted tasks, which may not reflect their everyday movement capabilities and can be challenging for individuals with advanced PD. There's a clear need for an objective, continuous, and less intrusive method to assess PD severity that can capture real-world movement patterns.

Methods

The core of this paper is the development of a Symbolic Movement Representation (SMR). The process involves:

  1. Data Acquisition: Continuous movement data is collected from wearable sensors (accelerometers on wrists).
  2. Filtering and Windowing: Raw sensor data is down-sampled, band-pass filtered, and segmented into 1-second windows with a 0.24-second overlap.
  3. Syllable Discovery (Clustering): A K-Means clustering algorithm is used to identify recurring patterns or "syllables" within these windows. The paper uses a vocabulary of 24 syllables, determined via an elbow method.
  4. Sequence Embedding: Each window of movement is then represented by its corresponding syllable, transforming the continuous time series into a sequence of symbols.
  5. Markov Chain Modeling: These sequences of syllables are modeled using a discrete-time Markov Chain (MC). The transition rates between syllables are estimated using maximum likelihood.
  6. Stationary Distribution (SMR): The stationary distribution of this Markov Chain is calculated. This distribution, the SMR, represents the statistical grammar of movement for a given session and is hypothesized to reflect the underlying neurological state. A higher degree of disorganization in syllable transitions (and thus a more dispersed SMR) is associated with increased motor impairment in PD.
  7. Distance Metric: The L1 norm (total variation distance) is used to compare SMRs between different sessions or participants.
  8. Dimensionality Reduction (MDS): Multi-Dimensional Scaling (MDS) is employed to visualize the relationships between SMRs in a lower-dimensional space, allowing for the assessment of how well SMR distances correlate with clinical scores.

The novelty lies in the unsupervised discovery of movement "syllables" and their statistical grammar (SMR) as a direct, unbiased measure of motor state, applicable to unconstrained movements.

Datasets

The study utilizes data from three sources, collectively referred to as the "Bluesky project":

  • In-clinic PD participants: 35 individuals with idiopathic PD were recruited and tested at Tufts Medical Center. Data was collected during two visits: one in the "ON" state (medication effective) and one in the "OFF" state (medication worn off).
  • Healthy participants (in-clinic): Data from 45 healthy individuals was collected at IBM and Pfizer sites. This group served as a baseline for comparison.
  • At-home PD participants: Data from 26 individuals with PD was collected continuously over six days (two 3-day periods separated by 30 days) using wearable devices at home.

The paper does not introduce a new benchmark dataset but rather uses existing data collected under specific protocols to validate their method.

Key Insights and Ideas

  • Movement as a Language: The paper frames human movement as a form of language, composed of "syllables" (basic movement motifs) and a "grammar" (the statistical relationships between these syllables). This analogy proves powerful for characterizing motor states.
  • Disorganization as a Biomarker: The core hypothesis is that Parkinson's disease leads to increasingly disorganized transitions between movement syllables, which is reflected in the stationary distribution of the Markov Chain (SMR). This disorganization serves as an objective biomarker for PD severity.
  • Unbiased and Unsupervised: The method is "unbiased" because it doesn't rely on pre-defined movement categories or human interpretation of specific tasks. It's "unsupervised" because it discovers the movement syllables and their statistical patterns directly from the data.
  • In-Clinic vs. At-Home Equivalence: The SMR approach demonstrates effectiveness in both controlled in-clinic settings with scripted tasks and in unconstrained, everyday at-home environments, bridging a critical gap in continuous monitoring.
  • Continuous Monitoring Potential: The ability to generate stable SMRs with relatively short periods of data (e.g., less than 2 minutes of aggregated ADLs) suggests the potential for highly granular, continuous monitoring of PD progression and treatment response.

Key Results

  • Correlation with Clinical Scores: The SMR, when analyzed using MDS, showed a strong correlation with the neurologist's assessment of motor impairment, particularly the Postural Instability and Gait Disorder (PIGD) subscore. The model derived from MDS components could estimate the UPDRS-III score with an R-squared of 0.64.
  • Distinguishing PD from Healthy: The SMR analysis clearly separated PD participants from healthy controls, with PD participants exhibiting greater dissimilarity in their SMRs, indicative of increased motor disorder.
  • ON/OFF State Discrimination: The method was able to distinguish between ON and OFF medication states in PD patients, suggesting sensitivity to dopaminergic fluctuations.
  • At-Home Validation: The SMR approach was successfully applied to continuous, at-home data, showing distinct diurnal and nocturnal movement patterns that correlated with overall UPDRS-III scores.
  • Inter-rater Variability: The model's estimation accuracy was comparable to the inter-rater variability observed among neurologists, highlighting its objectivity.

Limitations and Future Work

  • Sensor Placement: While the study primarily used wrist sensors, the paper notes that sensor placement can influence results, and further analysis on dominant vs. non-dominant hands and most/least affected hands was conducted. The best correlations were found for gait-specific impairments when using gait-specific data.
  • Tremor Exclusion: The current model, particularly when focusing on gait, might not capture tremor-related frequencies effectively, as tremor was largely excluded from the analysis to focus on other motor aspects.
  • Generalizability: While the method was tested on PD, the authors suggest it could be generalizable to detecting other neurological states with characteristic movement signatures.
  • Age Effect: While not a strong confounder, a weak correlation between age and MDS scores was observed in PD patients, suggesting age might have a minor influence.
  • Future Work: The authors suggest exploring the method's applicability to other neurological conditions and further refining the syllable discovery and Markov chain modeling for even greater precision. The potential for continuous, granular monitoring for therapy response and early detection is a significant avenue for future research.

Summary updated

Abstract

Unconstrained human movement can be broken down into a series of stereotyped motifs or 'syllables' in an unsupervised fashion. Sequences of these syllables can be represented by symbols and characterized by a statistical grammar which varies with external situational context and internal neurological state. By first constructing a Markov chain from the transitions between these syllables then calculating the stationary distribution of this chain, we estimate the overall severity of Parkinson's symptoms by capturing the increasingly disorganized transitions between syllables as motor impairment increases. Comparing stationary distributions of movement syllables has several advantages over traditional neurologist administered in-clinic assessments. This technique can be used on unconstrained at-home behavior as well as scripted in-clinic exercises, it avoids differences across human evaluators, and can be used continuously without requiring scripted tasks be performed. We demonstrate the effectiveness of this technique using movement data captured with commercially available wrist worn sensors in 35 participants with Parkinson's disease in-clinic and 25 participants monitored at home.
BibTeX
@inproceedings{abrami2020using,
  author = {Avner Abrami and Stephen Heisig and Vesper Ramos and Kevin Thomas and Bryan Ho and Vittorio Caggiano},
  title = {{Using an unbiased symbolic movement representation to characterize Parkinson’s disease states}},
  booktitle = {Scientific},
  year = {2020},
  doi = {10.1038/s41598-020-64181-3},
}
Figure