Electrodiagnostic accuracy in polyneuropathies: supervised learning algorithms as a tool for practitioners

Objective: The interpretation of electrophysiological findings may lead to misdiagnosis in polyneuropathies. We investigated the electrodiagnostic accuracy of three supervised learning algorithms (SLAs): shrinkage discriminant analysis, multinomial logistic regression, and support vector machine (SVM), and three expert and three trainee neurophysiologists. Methods: We enrolled 434 subjects with the following diagnoses: chronic inflammatory demyelinating polyneuropathy (99), Charcot-Marie-Tooth disease type 1A (124), hereditary neuropathy with liability to pressure palsy (46), diabetic polyneuropathy (67), and controls (98). In each diagnostic class, 90% of subjects were used as training set for SLAs to establish the best performing SLA by tenfold cross validation procedure and 10% of subjects were employed as test set. Performance indicators were accuracy, precision, sensitivity, and specificity. Results: SVM showed the highest overall diagnostic accuracy both in training and test sets (90.5 and 93.2%) and ranked first in a multidimensional comparison analysis. Overall accuracy of neurophysiologists ranged from 54.5 to 81.8%. Conclusions: This proof of principle study shows that SVM provides a high electrodiagnostic accuracy in polyneuropathies. We suggest that the use of SLAs in electrodiagnosis should be exploited to possibly provide a diagnostic support system especially helpful for the less experienced practitioners.