Machine learning-based skin nerve morphometry for diabetic neuropathy: diagnostic and clinical implications

This study aimed to (i) develop and validate new intraepidermal nerve fibres (IENFs) biomarkers with the aid of machine learning algorithms for the diagnosis of small-fibre neuropathy in diabetic patients and (ii) explore the diagnostic performance and clinical significance of these new biomarkers. Patients with diabetic neuropathy and control subjects were recruited. Area-based morphometry of IENF (IENFa) parameters were developed by using the machine learning system for automatic quantification. The diagnostic performance was assessed according to receiver operating characteristic analysis. The clinical implications of the various IENFa parameters were examined by exploring their correlations with metabolic profiles and via electrophysiological experiments. The diabetic neuropathy (n = 48) and control (n = 63) cohorts were comparable in terms of age and sex. The IENFa parameters were inversely correlated with age, and only the IENF density (IENFd, the number of fibres per unit length of epidermis) and IENFa/A parameters were observed to be sex dependent in the control group. All of the IENFa parameters demonstrated equivalent performance according to (i) the correlation with IENFd and (ii) the diagnosis of IENFd-based small-fibre neuropathy by the receiver operating characteristic analysis (area under curve: 0.91-0.95, P > 0.05). Furthermore, the IENFa biomarkers were significantly correlated with sural sensory nerve action potential amplitudes. In summary, automatic IENFa is time-efficient and performs comparably to IENFd in diagnosing diabetic small-fibre neuropathy with high reliability. Furthermore, the IENFa parameter reflects concurrent large-fibre involvement in diabetic neuropathy. As the IENFa represents the total area of all IENFs, the results also imply global axonal atrophy in diabetic neuropathy.