A Random Forest Classification Model for Spasticity after Stroke

Research Objectives: The objective of this analysis was to developed a Random Forest Classification Model to determine which sociodemographic and clinical health characteristics predict the development of post-stroke spasticity (PSS).

Design: Random Forest Classification Modelling with data collected from a retrospective review of patients’ medical charts and the National Rehabilitation Reporting System.

Setting: Specialized stroke rehabilitation inpatient and outpatient units from a large, urban centre in Ontario, Canada.

Participants: A total of 922 individuals (102 spastic) attended stroke rehabilitation between July 1, 2015 and June 30, 2020. A synthetic minority oversampling technique was used to generate a dataset of 1632 individuals (408 spastic).

Interventions: None

Main Outcome Measures: Clinical health data (i.e., number, type and location of stroke, age at stroke), length of stay, LOS); presence of spasticity, presence of aphasia, endovascular or tissue plasminogen activator treatment, Chedoke-McMaster impairment inventory scores, Functional Independence Measure motor and cognitive sub-scores.

Results: Based on the results of a feature hyper-parameterization procedure, a model with 1501 trees was selected with 8 variables (mtry=8). A random forest model was then trained using these parameters on 80% of the dataset. The relative importance of predictor variables was assessed by examining the decrease in model accuracy when a particular variable is permutated. Chedoke McMaster impairment scores, age at stroke onset, Functional Independence Measure motor sub-scores and length of stay appear to be the most important predictor variables when compared to the remaining 7. When testing the model performance on the untrained 20% of the dataset, the model performed with a balanced accuracy of 82.5%, a Cohen’s kappa value of 0.71 and a spasticity classification error of 32.1%.

Conclusions: Overall, our model performed relatively well, but could be tuned more for sensitivity to prioritize the correct identification of spasticity. In the future, more data will be required to validate the generalizability of the model and robustness of fit, as well as to eliminate the need for synthetically created data.

Author(s) Disclosures: None.