Reproducibility of In Vivo Corneal Confocal Microscopy Using an Automated Analysis Program for Detection of Diabetic Sensorimotor Polyneuropathy

OBJECTIVE In vivo Corneal Confocal Microscopy (IVCCM) is a validated, non-invasive test for diabetic sensorimotor polyneuropathy (DSP) detection, but its utility is limited by the image analysis time and expertise required. We aimed to determine the inter- and intra-observer reproducibility of a novel automated analysis program compared to manual analysis. METHODS In a cross-sectional diagnostic study, 20 non-diabetes controls (mean age 41.4±17.3y, HbA1c 5.5±0.4%) and 26 participants with type 1 diabetes (42.8±16.9y, 8.0±1.9%) underwent two separate IVCCM examinations by one observer and a third by an independent observer. Along with nerve density and branch density, corneal nerve fibre length (CNFL) was obtained by manual analysis (CNFLMANUAL), a protocol in which images were manually selected for automated analysis (CNFLSEMI-AUTOMATED), and one in which selection and analysis were performed electronically (CNFLFULLY-AUTOMATED). Reproducibility of each protocol was determined using intraclass correlation coefficients (ICC) and, as a secondary objective, the method of Bland and Altman was used to explore agreement between protocols. RESULTS Mean CNFLManual was 16.7±4.0, 13.9±4.2 mm/mm2 for non-diabetes controls and diabetes participants, while CNFLSemi-Automated was 10.2±3.3, 8.6±3.0 mm/mm2 and CNFLFully-Automated was 12.5±2.8, 10.9 ± 2.9 mm/mm2. Inter-observer ICC and 95% confidence intervals (95%CI) were 0.73(0.56, 0.84), 0.75(0.59, 0.85), and 0.78(0.63, 0.87), respectively (p = NS for all comparisons). Intra-observer ICC and 95%CI were 0.72(0.55, 0.83), 0.74(0.57, 0.85), and 0.84(0.73, 0.91), respectively (p<0.05 for CNFLFully-Automated compared to others). The other IVCCM parameters had substantially lower ICC compared to those for CNFL. CNFLSemi-Automated and CNFLFully-Automated underestimated CNFLManual by mean and 95%CI of 35.1(-4.5, 67.5)% and 21.0(-21.6, 46.1)%, respectively. CONCLUSIONS Despite an apparent measurement (underestimation) bias in comparison to the manual strategy of image analysis, fully-automated analysis preserves CNFL reproducibility. Future work must determine the diagnostic thresholds specific to the fully-automated measure of CNFL.