Corneal confocal microscopy for identification of diabetic sensorimotor polyneuropathy: a pooled multinational consortium study

Aims/hypothesis Small cohort studies raise the hypothesis that corneal nerve abnormalities (including corneal nerve fibre length [CNFL]) are valid non-invasive imaging endpoints for diabetic sensorimotor polyneuropathy (DSP). We aimed to establish concurrent validity and diagnostic thresholds in a large cohort of participants with and without DSP. Methods Nine hundred and ninety-eight participants from five centres (516 with type 1 diabetes and 482 with type 2 diabetes) underwent CNFL quantification and clinical and electrophysiological examination. AUC and diagnostic thresholds were derived and validated in randomly selected samples using receiver operating characteristic analysis. Sensitivity analyses included latent class models to address the issue of imperfect reference standard. Results Type 1 and type 2 diabetes subcohorts had mean age of 42 ± 19 and 62 ± 10 years, diabetes duration 21 ± 15 and 12 ± 9 years and DSP prevalence of 31% and 53%, respectively. Derivation AUC for CNFL was 0.77 in type 1 diabetes (p < 0.001) and 0.68 in type 2 diabetes (p < 0.001) and was approximately reproduced in validation sets. The optimal threshold for automated CNFL was 12.5 mm/mm2 in type 1 diabetes and 12.3 mm/mm2 in type 2 diabetes. In the total cohort, a lower threshold value below 8.6 mm/mm2 to rule in DSP and an upper value of 15.3 mm/mm2 to rule out DSP were associated with 88% specificity and 88% sensitivity. Conclusions/interpretation We established the diagnostic validity and common diagnostic thresholds for CNFL in type 1 and type 2 diabetes. Further research must determine to what extent CNFL can be deployed in clinical practice and in clinical trials assessing the efficacy of disease-modifying therapies for DSP. Electronic supplementary material The online version of this article (10.1007/s00125-018-4653-8) contains peer-reviewed but unedited supplementary material, which is available to authorised users.


Study Population
The study population was accrued from the baseline evaluations of a consortium of 5 different cohort studies initiated between 2008 and 2011. [1][2][3][4][5] [4] and Manchester, [3] cohorts and the Brisbane site has published other outcomes using baseline data [5]. This current study includes patient-level data from all baseline visits from all five sites, some of which was accrued after the publication of these preliminary works. In September 2014, the NIH funded continued longitudinal follow-up of participants from the studies initiated at each of these sites; this manuscript is the first to be published by this collaboration.
The presence of diabetes mellitus was defined in accordance with American Diabetes Association guidelines. Neuropathy due to non-diabetic causes, current eye infection or other conditions that precluded IVCCM, or allergy to the ocular anesthetic used during the IVCCM exam were exclusions. Neuropathies due to non-diabetic causes were determined through detailed patient history or through screening of immunoglobulins and B12 levels, depending on each site's local protocol. The protocol and consent procedures at all sites were approved by local research ethics boards, and written informed consent was provided by all study participants or their legal guardians.

Index Test (IVCCM Examination)
Participants underwent examination of the sub-basal nerve plexus of the cornea using the Heidelberg Tomograph Rostock Cornea Module III (Heidelberg Engineering GmbH, Heidelberg, Germany and Heidelberg Engineering, Smithfield, RI, USA) according to published methods. [6] In brief, topical anaesthetic and a viscous gel medium were applied, permitting a visual gel bridge between the cornea and the sterile single-use cap on the microscope's objective lens.
Images were taken through the sub-basal layer over a depth of 50 microns using methods that had minor procedural variation between centres. [7] The most technically sound images were identified manually by site staff, and IVCCM parameters were measured using a manual protocol and an automated protocol [8,9] that served as a method of standardization. The latter was also performed as it represents a significant resource-sparing tool that has not been systematically studied for validity compared to the standard manual method. For the manual protocol of image analysis, the examiner traced nerve fibres on the images using a graphic tablet and pen and the parameters were determined using semi-automated analytical software (CCM Image Analysis tool v0.6, developed by M. Dabbah, University of Manchester). For the automated protocol, fully-automated software determined the parameters (ACCMetrics Image Analysis Software v2.0, developed by M. Dabbah and X. Chen, University of Manchester). Results from 1-8 images per eye were averaged. Measured parameters were corneal nerve fibre length (CNFL), expressed as the total length of nerves in mm/mm 2 of image area; corneal nerve branch density (CNBD), expressed as the number of branches/mm 2 ; and corneal nerve fibre density (CNFD), expressed as the number of fibres/mm 2 . The automated protocol is known to provide measures of CNFL that are systematically 30% lower than manually-derived CNFL. [8,9] Subscripts AUTO and MANUAL indicate automated and manual quantification. Raters were either trained in optometry or ophthalmology, except Toronto, which used research assistants who underwent two-day training by the microscope manufacturer. Published data have demonstrated similar cohort IVCCM characteristics, reproducibility, and concurrent validity regardless of study centre. [1,2,4,7,[10][11][12]

Reference Standard (Clinical Evaluation and Nerve Conduction Studies)
All study participants underwent nerve conduction studies and comprehensive physical examination. For nerve conduction studies, all investigational sites measured the dominant limb peroneal and sural nerves using clinical nerve conduction study equipment according to the standards of the American Association for Neuromuscular and Electrodiagnostic Medicine. Each centre performed examinations and collected data independently and results were sent to the centre leading statistical analysis (Toronto) where an algorithm was used to determine neuropathy cases. Locally, the results of the neurological examinations were organized into clinical symptom and clinical sign scores, with different centres using different scores according to their baseline study protocols (which were determined prior to the formation of the current consortium). The scores for symptoms included neuropathy symptom score (NSS), neuropathy symptom profile (NSP), and diabetic neuropathy symptom (DNS) score and the scores for signs included the neuropathy disability score (NDS) and Toronto clinical neuropathy score (TCNS).
The algorithm for the reference standard was positive if two criteria were met: 1) if a clinical symptom or clinical signs were present and 2) if peroneal motor nerve conduction velocity was less than 42 m/s. This reference standard was based on consensus criteria. [13][14][15] Reference standard definitions using other combinations of abnormal peroneal and/or sural nerve parameters and/or presence of signs and symptoms were considered in sensitivity analysis.
Other nerve conduction parameters included sural nerve amplitude potential and conduction velocity, and peroneal nerve amplitude potential, conduction velocity, and F-wave latency (adjusted for height). Amplitude potentials were adjusted for age. Specifically, these alternate reference standards included the following 6 case-definitions: i) abnormal sural nerve amplitude potential (≤7.2 µV if age ≤65 and ≤5.5 µV if age >65); ii) using each centre's local casedefinition where nerve conduction values were considered abnormal when greater than the 99th percentile or less than the 1st percentile in the reference healthy population database used at each site, according to the ranges determined during certification of the clinical laboratories; iii) any abnormal sural nerve parameter, corroborated by any abnormal peroneal parameter, corroborated by the presence of at least one sign or one symptom ("Toronto site's definition"); iv) the study's reference standard described above; v) the presence of a high number of signs and/or symptoms ("Clinical Definition"); and vi) the presence of a high number of signs and/or symptoms, corroborated by peroneal motor nerve conduction velocity <42 m/s ("Stringent Clinical Definition").

Variables Used for Sensitivity Analyses
Sensitivity analyses were undertaken to account for an imperfect reference standard. [14] Specifically, electrophysiology and clinically relevant signs and symptoms detect large nerve fibre dysfunction and may fail to identify patients with early small fibre neuropathy detected using IVCCM. [16] Skin Biopsy for IENFD assessment was undertaken in a subset of the Manchester participants, [17] and was used in resolver test analysis (Sensitivity Analysis #2).
Other evaluations included blood pressure, smoking history and biochemical tests including glycated hemoglobin A 1c , serum lipids, and urinary albumin excretion (generally conducted on the same day as or within 1 week of the neuropathy evaluation). Cooling detection thresholds were determined in the majority of participants by CASE IV (WR Medical Electronics Co., MN, USA) or the Medoc TSA-II NeuroSensory Analyzer (Medoc Advanced Medical Systems, Ramat-Yishai, Israel), using the method of limits.

Statistics
The available sample size yielded a power of >0.99 to detect a conservative area under the receiver operating characteristic curve (AUC) of 0.70 from the null hypothesis of 0.50. AUC was compared within study populations using the method of Pencina et al. [18] AUC was compared between study populations using the method of Hanley & McNeil. [19] Optimal diagnostic thresholds were identified by distance to the point of perfect discrimination using the formula The sensitivity analyses to account for imperfect reference standard included: 1) modification of the reference standard definition parameters to create less-and more-stringent definitions (as described above); 2) use of composite reference standard methods incorporating small fibre measures of intra-epidermal nerve fibre density into the definition of neuropathy cases; [17] and 3) latent class analysis that identified clusters of patients who shared common clinical characteristics and neurological test results (including the clinical scales, cooling detection threshold tests, [20] and electrophysiological tests) that were consistent with the presence of neuropathy. [21] Details of variables used in these analyses are provided above. ROC regression was performed according to the method of Janes, Longton, and Pepe. [22] Additionally, while we focused on the optimal diagnostic threshold as the single value that simultaneously maximized sensitivity and specificity, we conducted an alternate approach in which a pair of diagnostic thresholds were determined -one lower value chosen to maximize specificity (more confidently rule in the presence of neuropathy) and one higher value chosen to simultaneously maximize sensitivity (more confidently rule out the presence of neuropathy). To accomplish this, we used a combination of decision criteria that included threshold values that i) maximized positive and negative likelihood ratios, and ii) minimized false positives and false negatives.

ESM Tables
ESM Table 1. Characteristics of the 998 study participants, according to study centre.

Characteristic
Total N=998 Brisbane n=235 Calgary n=84 Manchester n=281 Michigan n=17 Due to differences in case-composition between definitions for DSP, comparisons of AUC were made using the method of Hanley and McNeil [19]. CNFL, corneal nerve fibre length; DM, diabetes mellitus; T1DM, type 1 diabetes mellitus; T2DM, type 2 diabetes mellitus.
ESM Fig. 1 Legend: Index test was IVCCM and reference standard was clinical and electrophysiological assessment. T1DM, type 1 diabetes participants; T2DM, type 2 diabetes participants.
ESM Fig. 2 Legend: Definitions to the left of the study definition represent less stringent definitions of neuropathy, while definitions to the right represent more stringent definitions. Red symbols indicate AUC for the study's reference standard definition. The solid horizontal lines represent the AUC for CNFL AUTO using the study definition as the reference standard. The dashed horizontal lines represent the AUC for CNFL MANUAL using the study definition as the reference standard. Error bars represent the standard error for AUC. CNFL, corneal nerve fibre length; DSP, diabetic sensorimotor polyneuropathy; T1DM, type 1 diabetes mellitus; T2DM, type 2 diabetes mellitus; Def n , definition.