Agreement between automated and manual quantification of corneal nerve fiber length: Implications for diabetic...

Quantification of corneal nerve fiber length (CNFL) by in vivo corneal confocal microscopy represents a promising diabetic neuropathy biomarker, but applicability is limited by resource-intensive image analysis. We aimed to evaluate, in cross-sectional analysis of non-diabetic controls and patients with type 1 and type 2 diabetes with and without neuropathy, the agreement between manual and automated analysis protocols.

Sixty-eight controls, 139 type 1 diabetes, and 249 type 2 diabetes participants underwent CNFL measurement (N=456). Neuropathy status was determined by clinical and electrophysiological criteria. CNFL was determined by manual (CNFL_Manual, reference standard) and automated (CNFL_Auto) protocols, and results were compared for correlation and agreement using Spearman coefficients and the method of Bland and Altman (CNFL_Manual subtracted from CNFL_Auto).

Participants demonstrated broad variability in clinical characteristics associated with neuropathy. The mean age, diabetes duration, and HbA1c were 53±18years, 15.9±12.6years, and 7.4±1.7%, respectively, and 218 (56%) individuals with diabetes had neuropathy. Mean CNFL_Manual was 15.1±4.9mm/mm², and mean CNFL_Auto was 10.5±3.7mm/mm² (CNFL_Auto underestimation bias, -4.6±2.6mm/mm² corresponding to -29±17%). Percent bias was similar across non-diabetic controls (-33±12%), type 1 (-30±20%), and type 2 diabetes (-28±16%) subgroups (ANOVA, p=0.068), and similarly in diabetes participants with and without neuropathy. Levels of CNFL_Auto and CNFL_Manual were both inversely associated with neuropathy status.

Although CNFL_Auto substantially underestimated CNFL_Manual, its bias was non-differential between diverse patient groups and its relationship with neuropathy status was preserved. Determination of diagnostic thresholds specific to CNFL_Auto should be pursued in diagnostic studies of diabetic neuropathy.