Innovative Model Predicts Upper Limb Capacity Recovery Profiles Poststroke


The researchers noted that the recovery model can use all available serially assessed data in a flexible way, creating a prediction at any desired moment poststroke, stand-alone, or linked with an electronic health record system.

Ruud Selles, PhD

An innovative dynamic model that used multiple measurements assessed at non-fixed time points helped researchers predict real-time, patient-specific upper limb capacity recovery profiles up to 6 months in patients with poststroke.1

A cohort of 450 patients with a first-ever ischemic hemispheric stroke had their recovery profile measured using the Action Research Arm Test (ARAT) for at least 3 assessment sessions, starting within the first week until 6 months poststroke. Lead author Ruud Selles, PhD, associate professor, Erasmus MC, and colleagues found that among the mixed-effect models used, a model with only ARAT time course, finger extension and shoulder abduction performed as good as models with more covariates.

"The model can use available serially assessed data in a flexible way, creating a prediction at any desired moment poststroke, stand-alone or linked with an electronic health record system,” Selles et al concluded.

The investigators performed a 5-fold cross-validation procedure by splitting the data into 5 sets, fitted the models using 4 of them, and then predicting the recovery profiles for the patients from the fifth set. More specifically:

  1. Model 1. ARAT scores as a function of all available covariates with their main effects and interaction with time.
  2. Model 2. ARAT scores as a function of all covariates with all their main effects but only the significant interactions with time.
  3. Model 3. ARAT scores as a function of only the significant main effects and the significant interactions with time.
  4. Model 4. ARAT scores as a function of time, the Shoulder-Adduction-Finger Extension (SAFE) model and their interaction with time.
  5. Model 5. ARAT scores as a function only of time.

Comparing the prediction accuracy of the different mixed-effect models and taking the first measurement of eat patient as a predictor, Selles and colleagues found mean cross-validation errors at 6 months poststroke of 10.1 to 105 ARAT points for the first, second and third model, compared to 8.4 for the fourth SAFE model with only finger extension (FE) and shoulder abduction, and 17.3 for the fifth model that includes only time as a covariate.

READ MORE: IND Submitted for ImmCelz Regenerative Stroke Immunotherapy

Using the SAFE model, increasing the number of serial measurements decreased the absolute error of prediction of ARAT from 8.4 points on the ARAT (Q1–Q3: 1.7–28.1) to 2.3 (Q1–Q3: 1–7.2) when 7 serial measurements were used for predicting the outcome at 6 months.

Prediction errors were relatively small at each time point throughout the 6-month observation period for patients with a higher initial ARAT score, while those with a low ARAT score saw large errors at baseline but decreased strongly later poststroke.

A subset of 2 patients using the SAFE model showed that predicted recovery is nonlinear, especially in the patient with low initial ARAT scores. Patient 1 had low ARAT scores at baseline and a moderate predicted ARAT score at 6 months, whereas Patient 2 had a higher ARAT score at baseline and the predicted score at 6 months was >50 points.

Investigators of the study included the link to access the real-time tool online, which includes instructions on how to enter data from individual patients and predict upper extremity recovery. For access to the Dynamic Prediction ARAT app, click here.

Selles et al wrote, “this is also the model that, within our own setting, is coupled to an online structured data entry system called profits, as a model of how more computationally complex prediction models can be coupled to electronic health record data or other types of online healthcare data collection systems.”

They also noted that future impact studies may indicate which level of detail is needed for clinical care and how knowledge of predicted outcome for individual influences impacts the decision making in stroke rehabilitation.

In July 2020, Peter McAllister, MD, medical director, New England Institute for Neurology and Headache, spoke about the lack awareness and options for those who suffer from upper limb spasticity following a stroke. Look back at the interview with NeurologyLive below.2

1. Selles R, Andrinopoulou ER, Nijland R, et al. Computerised patient-specific prediction of the recovery profile of upper limb capacity within stroke services: the next step. J Neurol Neurosurg Psychiatry.Published online January 21, 2021. doi: 10.1136/jnnp-2020-324637
2. Revance announces last patient enrolled in modified JUNIPER phase 2 upper limb spasticity trial of daxibotulinumtoxinA for injection [news release]. Newark, NJ: Revance. Published June 30, 2020. Accessed January 28, 2021.
Related Videos
Ro'ee Gilron, PhD
Monica Verduzco-Gutierrez, MD
Shahid Nimjee, MD, PhD
Peter J. McAllister, MD, FAAN
Video 6 - "Utilization of Neuroimaging in Alzheimer’s Disease"
Video 5 - "Contribution of Multiple Pathways to the Development of Alzheimer’s Disease"
Michael Levy, MD, PhD
Michael Levy, MD, PhD
© 2024 MJH Life Sciences

All rights reserved.