To determine how well nerve conduction study (NCS) data can be analyzed by recent Machine Learning (ML) algorithms to characterize patient demographics and biopsy evaluated Nerve Fiber Density (NFD).

The NCS best evaluates the fastest, largest nerve fibers which produce their signal early in nerve action potentials, but the signal from slow, small fibers is lost among other signals in the later waveform. Emerging ML algorithms, including Generative Adversarial Networks (GANs), have been used in signal, sound, and image analysis, but have not yet been applied to NCS data. Skin biopsies evaluating NFD are a gold standard for evaluating small fiber health.

This retrospective cohort pilot study evaluated patients seen over a 5 year period who received NCS evaluation followed by skin biopsies. The age, gender, biopsy results, and NCS data for 76 patients was used to train a GAN and neural network classifier to analyze NCS data. The model’s accuracy was judged from predictions on data not used to create the model.

Notional, artificial NCS data was generated from the model, which was qualitatively plausible compared to actual data. The model’s prediction of age from an NCS recording was poor (Average error 2.0 years, std 17.2). The model could use a single NCS recording to predict gender with no accuracy (c-statistic 0.51) and biopsy confirmed reduced NFD with some accuracy (c-statistic 0.59).

The health of small fiber nerves can be interrogated in NCS evaluation using advanced tools, however these tools will require continued study and prospective, normative validation before they can provide clinical utility. Emerging ML algorithms present exciting opportunities to improve neurologic diagnosis and research.