Abstract Details

Title Using resting-state electroencephalography to predict motor function improvement in spinal cord injury patients: a machine learning approach.

Topic Neuro-rehabilitation

Presentation(s) Neuro-rehabilitation Posters (7:00 AM-5:00 PM)

Poster/Presentation
Number 005

Objective We aimed to implement a machine learning (ML) algorithm to predict non-invasive brain stimulation effectiveness combined with robotic gait training for motor improvement in spinal cord injury (SCI) patients.

Background The transcranial direct current stimulation (tDCS) and robotic gait training have shown potential benefits for SCI motor rehabilitation, but the high variation of response among patients reduces their clinical translation; thus, a ML-based predictive model of response could help us identify good responder candidates to guide the resources allocation better.

Design/Methods We applied a supervised ML approach to develop a classification model to predict motor function response in SCI patients after tDCS combined with robotic gait training. The primary outcome was motor function improvement (indexed by WISCI-II scale). Demographical, clinical, and electroencephalography variables (absolute and relative delta, theta, alpha, and beta band power from frontal, central and parietal areas) were included as predictors. An ML approach of feature engineering, data pre-processing, and model optimization was used to create the most accurate predictive model. We performed a cross-validation process. We trained multiple ML classifiers: Support Vector Machine, Decision Trees, Naive Bayes, and Multilayer Perceptron (MLP). We used the AUC of ROC for models comparison. A feature importance analysis algorithm was performed to obtain the most differentiable features.

Results This analysis included 39 incomplete SCI patients. The best performing predictive model, used an MLP classifier and had an average area under the curve (AUC) of 0.81. The most important features in the model were high-beta power of parietal and fronto-central areas, and delta power of parietal and frontal areas. The intervention variable (tDCS group vs sham tDCS group) was an independent moderator of the classification accuracy.

Conclusions Our model predicted motor improvement after tDCS combined with robotic gait training with good accuracy. Our analysis suggests the applicability of machine-learning methods for individualized treatments in SCI rehabilitation.

Authors/Disclosures
Kevin Pacheco-Barrios, MD (Neuromodulation Center, Harvard Medical School) PRESENTER	Dr. Pacheco-Barrios has nothing to disclose.
Marcel Simis, MD, PhD (University of Sao Paulo)	No disclosure on file
	No disclosure on file
	No disclosure on file
	No disclosure on file
Felipe Fregni (Spaulding Rehabilitation Hospital)	Felipe Fregni has nothing to disclose.

��ɫ��

��ɫ��