Predicting and Monitoring Upper-Limb Rehabilitation Outcomes Using Clinical and Wearable Sensor Data in Brain Injury Survivors

Sunghoon Ivan Lee, Catherine Adans-Dester, Anne O’Brien, Gloria Vergara-Díaz, Randie M. Black‐Schaffer, Ross Zafonte, Jennifer Dy, Paolo Bonato

Abstract: Precision rehabilitation requires accurate prediction and monitoring of patient recovery. This study introduces a machine learning framework that integrates routine clinical assessments with wearable sensor data to estimate functional outcomes in stroke and traumatic brain injury (TBI) survivors, achieving human-level assessment accuracy.

Introduction

Rehabilitation is not a "one size fits all" process. Two patients with similar brain injuries can exhibit vastly different recovery trajectories, yet current clinical models often rely on broad baseline assessments like age or initial impairment levels. These models frequently fail to capture the high degree of individual variability in treatment response. This research addresses the dual challenge of predicting long-term improvement and objectively monitoring movement quality through a novel integration of clinical and wearable sensor data.

Methodology

The study utilized a multi-stage machine learning pipeline involving 44 subjects (22 stroke and 22 TBI survivors). To capture the "fine-grained" nuances of movement, participants wore five accelerometer units placed strategically on the sternum, paretic arm, wrist, thumb, and index finger while performing tasks from the Wolf Motor Function Test (WMFT).

Sensor placement on the sternum and paretic limb to capture movement quality metrics.

The core innovation of this research is the Aggregation Model. The team extracted over 1,000 features from the sensor data, focusing on movement quality metrics like speed, smoothness, and coordination. They then built two separate Gaussian Process Regression (GPR) models: one estimating the Functional Ability Scale (FAS) score using clinical variables, and another using only wearable sensor features.

The proposed machine learning framework for outcome estimation, combining clinical and sensor data.

Instead of simple averaging, the Aggregation Model uses a constrained linear model to determine how much to "trust" sensor data versus clinical data for a specific patient, effectively using clinical characteristics as an anchor when necessary.

Key Results and Insights

The performance of the combined model was evaluated against ground-truth scores provided by three trained clinicians. The results demonstrated a significant leap in predictive power:

  • Clinical Data Only: $r = 0.79$, RMSE = 10.8%.
  • Sensor Data Only: $r = 0.91$, RMSE = 7.7%.
  • Proposed Combined Model: $r = 0.94$, RMSE = 6.9%.

The combined model's error rate (6.9%) is remarkably close to the typical variability seen between different human clinicians, suggesting the algorithm is approaching human-level assessment accuracy.

Crucially, the researchers found that clinical data tends to predict the "average" responder, often regressing to the mean. In contrast, sensor-based models were far better at capturing individual variability and "Aha!" moments of improvement that clinical scores might miss. On average, the aggregation model placed 93% of the weight on sensor data, though clinical data remained a necessary component for certain patient profiles.

Discussion and Future Work

This unified approach performed similarly for both stroke and TBI survivors, suggesting that the underlying mechanics of motor recovery assessment are comparable across different brain injury types.

While the current results are promising, limitations include the relatively small sample size and the "sensor burden" of wearing five units. Future work will focus on identifying a "minimum set" of sensors (such as a single wristband) and applying these models to "free-living" data to monitor how patients use their limbs in their own homes during daily activities. Incorporating physiological data, such as MRI-based corticospinal tract integrity, could further refine these predictive capabilities.

Summary updated

Abstract

This is important in the context of developing precision rehabilitation interventions.
BibTeX
@article{lee2020predicting,
  author = {Sunghoon Ivan Lee and Catherine Adans-Dester and Anne O’Brien and Gloria Vergara-Díaz and Randie M. Black‐Schaffer and Ross Zafonte and Jennifer Dy and Paolo Bonato},
  title = {{Predicting and Monitoring Upper-Limb Rehabilitation Outcomes Using Clinical and Wearable Sensor Data in Brain Injury Survivors}},
  journal = {IEEE Transactions on Biomedical Engineering},
  year = {2020},
  doi = {10.1109/tbme.2020.3027853},
}
Figure