Poster abstracts

Poster number 46 submitted by Sarah Hulbert George

Predicting the extent of post-stroke motor recovery under two modalities of constraint-induced movement (CI) therapy using enhanced probabilistic neural networks (EPNN)

Sarah Hulbert George (Biophysics), Mohammed Hossein Rafiei (Civil Engieering), Lynne Gauthier (Physical Medicine and Rehabilitation), Alexandra Borstad (Physical Therapy, College of St. Scholastica), John Buford (Physical Therapy), Hojjat Adeli (Biophysics, Civil Engineering, and Neuroscience)

Abstract:
The ability to functionally move ones arm is often lost or diminished after stroke. Motor rehabilitation, and constraint-induced movement (CI) therapy in particular, is an effective method for increasing ones motor ability of the more affected upper extremity. However, the extent of motor recovery for individuals that undergo CI therapy can vary widely. One reason for this variation is because recovery depends on many factors such as the person’s baseline (post-stroke, pre-therapy) fine motor, gross motor, and somatosensory ability. In this study, we evaluated the predictive power of each of these factors in the extent of motor recovery after CI therapy using an enhanced probabilistic neural network (EPNN) prognostic model. These factors served as inputs to EPNN and were measured using the Wolf Motor Function Test (WMFT), the Brief Kinesthesia Test, and the Touch Test Monofilaments. We investigated these factors in two different modalities of CI therapy. The first modality was the standard CI therapy administered in the clinic, the second was a modified Kinect-based video game version of CI therapy1 that was administered primarily in the home. We found that, across therapy modalities, gross motor ability at baseline, as measured by the WFMT, seemed to be the most robust predictor of the extent of recovery, where somatosensory input was, at best, only moderately predictive. In the standard and gaming CI therapy data sets, we were able to achieve EPNN models with maximum prediction accuracies of 100% and 94.7%, respectively. In the combined model, we achieved maximum accuracies of 94.5%. Therefore, we were able to predict, with high accuracy, the extent of recovery of post-stroke participants given only their baseline scores and further, to determine which factors contained the highest predictive power. This improved predictability has the potential to help physicians make more personalized, cost-effective therapy decisions for their patients.

References:
[1] Liang, Jiongqian, et al. "Data Analytics Framework for A Game-based Rehabilitation System." Proceedings of the 6th International Conference on Digital Health Conference. ACM, 2016.

Keywords: Chronic Stroke, Enhanced Probabilistic Neural Networks, Motor Function