Enabling precision rehabilitation interventions using wearable sensors and machine learning to track motor recovery

Catherine P. Adans-Dester, Nicolas Hankov, Anne T. O'Brien, Gloria Vergara-Diaz, Randie Black-Schaffer, Ross Zafonte, Jennifer G. Dy, Sunghoon Ivan Lee, Paolo BonatoNpj Digit. Medicine 2020 · DOI · OpenReview

Introduction

Rehabilitation for neurological conditions, such as stroke and traumatic brain injury (TBI), is characterized by significant inter-patient variability. Two individuals with comparable initial injuries often exhibit vastly different responses to identical therapeutic interventions. To optimize recovery, clinicians require "precision rehabilitation"—the ability to tailor treatments to a patient's specific, evolving needs.

Currently, the data necessary for such precision is lacking. The gold standard for assessment, the Fugl-Meyer Assessment (FMA), is time-consuming and subjective, typically leading to data collection only at admission and discharge. This creates a "black box" during the treatment period, preventing therapists from identifying plateaus or the need for strategic shifts in real-time. This research addresses this gap by utilizing wearable sensors and machine learning to provide continuous, objective tracking of motor recovery.

Methods and Data Collection

The study involved 37 participants: 16 stroke survivors and 21 TBI survivors, all presenting with residual upper-limb hemiparesis. Despite the different etiologies, the researchers found no significant difference in movement patterns or rehabilitation responses between the groups, allowing for a combined cohort analysis.

Participants were equipped with six wearable accelerometers placed on the chest, upper arm, forearm, wrist, and fingers. They performed eight functional tasks from the Wolf Motor Function Test, such as lifting a can or turning a key.

Wearable sensor placement and activity recording during functional tasks.

Technical Approach

The core of the methodology is a multi-stage machine learning pipeline designed to translate raw sensor data into clinical scores.

  1. Feature Extraction: The system derives features from accelerometer time series, including signal entropy, jerk, and correlation coefficients between body segments. These features are specifically chosen to capture compensatory movements, such as trunk leaning during reaching tasks.
  2. Balanced Random Forest (RF): To address the non-uniform distribution of clinical scores in the dataset, the authors employed a "balanced" RF approach. By sub-sampling training data, the model learns effectively across all levels of impairment severity, reducing bias toward the most common scores.
  3. Cross-Domain Input: A key innovation is the use of estimated movement quality (Functional Ability Scale, or FAS) as an input feature to predict physiological impairment (FMA). This mimics clinical intuition, where a therapist's observation of movement quality informs their assessment of underlying neurological status.
Overview of the multi-stage machine learning pipeline for estimating clinical scores.

Key Results

The machine learning framework demonstrated high accuracy and strong agreement with clinician-rated scores:

  • Impairment (FMA): The model achieved a coefficient of determination ($r^2$) of 0.86 and a Root-Mean-Square Error (RMSE) of 3.99 points on a 66-point scale. This represents a significant improvement over baseline linear regression models ($r^2 = 0.47$).
  • Movement Quality (FAS): The system achieved an $r^2$ of 0.79 and an RMSE of 0.38 points on a 5-point scale.
  • Bias Reduction: The balanced training approach successfully mitigated the tendency to overestimate scores for severely impaired patients.

Discussion and Future Directions

A significant insight from this work is the deep link between movement quality and physiological impairment. Furthermore, the study highlights the utility of Random Forest architectures in handling high-dimensional sensor data within relatively small clinical datasets.

The ability to track recovery longitudinally allows clinicians to compare patient progress against theoretical models, such as the "proportional restitution" or "70% rule," which suggests many patients recover approximately 70% of their initial impairment. Early identification of "non-responders" could trigger timely adjustments to therapy.

Future work will focus on transitioning from task-based recording to "passive monitoring" in home environments and expanding the cohort size. Additionally, the researchers aim to move from estimation to predictive modeling, using longitudinal data to forecast long-term recovery outcomes.

Summary updated

Abstract

The need to develop patient-specific interventions is apparent when one considers that clinical studies often report satisfactory motor gains only in a portion of participants. This observation provides the foundation for “precision rehabilitation”. Tracking and predicting outcomes defining the recovery trajectory is key in this context. Data collected using wearable sensors provide clinicians with the opportunity to do so with little burden on clinicians and patients. The approach proposed in this paper relies on machine learning-based algorithms to derive clinical score estimates from wearable sensor data collected during functional motor tasks. Sensor-based score estimates showed strong agreement with those generated by clinicians. Score estimates of upper-limb impairment severity and movement quality were marked by a coefficient of determination of 0.86 and 0.79, respectively. The application of the proposed approach to monitoring patients’ responsiveness to rehabilitation is expected to contribute to the development of patient-specific interventions, aiming to maximize motor gains.
BibTeX
@article{adansdester2020enabling,
  author = {Catherine P. Adans-Dester and Nicolas Hankov and Anne T. O'Brien and Gloria Vergara-Diaz and Randie Black-Schaffer and Ross Zafonte and Jennifer G. Dy and Sunghoon Ivan Lee and Paolo Bonato},
  title = {{Enabling precision rehabilitation interventions using wearable sensors and machine learning to track motor recovery}},
  journal = {Npj Digit. Medicine},
  year = {2020},
  doi = {10.1038/s41746-020-00328-w},
}
Figure