43. The quest for optimal electrodiagnosis of Guillain–Barre’ syndrome subtypes: criteria sets versus a linear discriminant analysis model

Abstract

Up to 38% of Guillain-Barré syndrome (GBS) patients change electrophysiological classification after serial recordings. Aim of this study is to compare at the first electrodiagnostic test the predictivity of final diagnosis of three diagnostic criteria sets versus a regularized form of linear discriminant analysis (LDA) model. In 98 GBS patients with serial conduction studies in three motor and sensory nerves 24 electrophysiological variables were analyzed. Final diagnosis, established by serial studies and result of anti-ganglioside antibodies testing, revealed 54 demyelinating and 44 axonal GBS. We applied a sparse LDA method which is a technique for partitioning a vector of parameters (predictors) into one of the two classes (demyelinating or axonal) based on a linear projection learned from the 98 labeled patients. The evaluation of the discriminant power of sparse LDA is based on the leave one out cross validation procedure. We also employed the electrodiagnostic criteria sets developed by Hadden et al. (1998), by Rajabally et al. (2015) and by us. The overall misclassification error rate of sparse LDA at first test is 14% with a sensitivity for demyelinating GBS of 87% and a specificity of 84%. Distal motor latency, duration of distal motor potentials, proximal/distal motor amplitude ratio and sensory nerve action potential amplitude play a prominent role for the appropriate classification of patients. The overall misclassification error rates with the electrodiagnostic criteria are: 45% for Hadden’s, 43% for Rajabally’s and 26% for ours. Conclusions: (1) the model based on sparse LDA performs much better than the “a priori” established criteria sets; (2) the criteria set we developed, including the analysis of duration of distal motor potentials and of sensory nerve action potential amplitude, performs better than the other two; (3) only serial studies were able to demonstrate reversible conduction failure in motor and sensory fibers.