Study design and study population

The German Diabetes Study (GDS) is an ongoing observational prospective study that evaluates the natural course of recently diagnosed diabetes and explores prognostic factors and mechanisms leading to diabetes-related complications [31]. Individuals aged 18–69 years at the baseline examination with a known diabetes duration of less than 1 year are eligible to participate, while individuals with overt neurological diseases, such as multiple sclerosis, dementia or psychiatric disorders, are excluded. Diabetes is diagnosed according to the ADA criteria [32]. At baseline, all participants undergo a comprehensive examination consisting of clinical tests, a face-to-face interview, standardised written questionnaires and detailed laboratory measurements.

The GDS is conducted according to the Declaration of Helsinki, approved by the ethics committee of Heinrich Heine University, Düsseldorf, Germany (ref. 4508) and was registered with ClinicalTrials.gov (registration no. NCT01055093). All participants provided written informed consent.

This cross-sectional analysis focused on 503 participants recruited consecutively from 2005 to 2011. From these 503 participants, we excluded 51 participants with missing data on one of the neurological variables used to define DSPN, two participants with diabetes forms other than type 1 or type 2 diabetes and 36 participants with missing covariates, leaving 423 participants with complete data for analysis (ESM Fig. 1).

Measurement of serum NFL

NFL was measured in serum of fasting participants at baseline using the Olink Target 96 Neuro Exploratory multiplex assay panel (Olink Proteomics, Uppsala, Sweden; NFL Uniprot ID: P07196 and NFL Olink ID: OID05206) based on proximity extension assay technology. The relative concentration of serum NFL is expressed as normalised protein expression (NPX) values, which are comparable in their distribution to log2-transformed protein concentrations. A detailed description of all 92 biomarkers in the panel is given in ESM Table 1; data cleaning steps that left 60 biomarkers, in addition to NFL, for exploratory analysis are described in ESM Fig. 2. We first excluded 29 biomarkers because of missing data for ≥25%. Then, we further excluded two biomarkers because of inter-assay CV >25%.

Assessment of peripheral neuropathy

All participants underwent nerve conduction studies (NCSs) and quantitative sensory testing (QST) as previously described [33, 34]. Motor nerve conduction velocity (MNCV) was measured in the peroneal, median and ulnar nerves, while sensory nerve conduction velocity (SNCV) and sensory nerve action potential (SNAP) were measured in the sural, median and ulnar nerves. All nerve stimulations were performed using surface electrodes (Nicolet VikingQuest; Natus Medical, San Carlos, CA, USA) after warming up feet and lower legs to ensure that skin temperature was 33–34°C. QST was evaluated by thermal detection thresholds (TDTs) to warm and cold stimuli at the thenar eminence and dorsum of the foot using the method of limits (TSA-II NeuroSensory Analyzer; Medoc, Ramat Yishai, Israel). Neurological examination was performed using the Neuropathy Disability Score (NDS) for neuropathic signs and the Neuropathy Symptom Score (NSS) for neuropathic symptoms [35]. Stages of DSPN were defined, according to the Toronto Consensus criteria [36], as subclinical, confirmed asymptomatic and confirmed symptomatic as previously described [33].

Covariates

Body weight (kg), waist circumference (cm) and height (m) were measured at enrolment. Information on known diabetes duration (days), presence of chronic diseases, use of non-steroidal anti-inflammatory drugs (NSAIDs) (yes/no) and lipid-lowering drugs (yes/no) was obtained during a face-to-face interview. Hypertension was defined as either systolic BP ≥140 mmHg, diastolic BP ≥90 mmHg or use of antihypertensive medication. Self-reported myocardial infarction, peripheral arterial occlusive disease, cerebrovascular disease or stroke was used to define the presence of CVD. Total cholesterol and HbA1c were measured according to standardised laboratory procedures in blood samples collected at baseline after overnight fasting [31]. Diabetes-related autoantibodies were measured for each participant. The eGFR was calculated from serum creatinine and cystatin C using the Chronic Kidney Disease Epidemiology (CKD-EPI) equation [37].

Statistical analyses

Data are presented as median (25th, 75th percentiles), mean ± SD, or percentages in descriptive statistics. Differences in characteristics according to DSPN status were tested with generalised linear regression analyses allowing for different group variances using the SAS procedure GLIMMIX and with the χ2 test, while differences in serum NFL levels between the two groups were tested with ANCOVA to allow adjustment for age at diagnosis. In addition, differences in means and 95% CIs were also calculated. Spearman’s rank correlation tests estimated the non-adjusted and age-adjusted correlations of serum NFL with demographic and metabolic variables.

In primary analyses, we assessed the associations of serum NFL with DSPN (primary endpoint) and with nerve function measures (exploratory secondary endpoints). First, the association of serum NFL (independent variable, by 1-NPX increase, by 1-SD increase and by tertiles) with DSPN (binary dependent variable) was assessed using Poisson regression models with a robust error variance. Model 1 was adjusted for sex and age at diagnosis. Model 2 was additionally adjusted for waist circumference, height, HbA1c, known diabetes duration, diabetes type, eGFR, total cholesterol, hypertension, CVD, lipid-lowering drugs and NSAIDs. Based on our previous studies, these covariates were defined a priori as covariates that may affect nerve conduction [38, 39]. Associations were estimated with RRs of DSPN and their corresponding 95% CIs. Receiver operating characteristic (ROC) curves were used to assess the predictive performance of serum NFL. We conducted the following sensitivity analyses to test the robustness of our results: (1) we substituted waist circumference and height with BMI; (2) we excluded individuals with type 1 diabetes; and (3) we excluded participants with prevalent self-reported CVD (n=10). In addition, we tested for interaction with diabetes type. Second, the associations of serum NFL with MNCVs and SNCVs were assessed using multivariable linear regression models, while associations with SNAPs and TDTs were assessed using quantile regression models, which do not make assumptions about the residual distribution. These analyses were adjusted for the same covariates as in the Poisson regression models, and associations were estimated with β estimates and their corresponding 95% CIs. In addition, for each participant, individual MNCVs, SNCVs and SNAPs were standardised and summarised in sum scores, which were used as additional secondary outcomes analysed using multivariable linear regression models. Z scores were calculated by subtracting the mean value of nerve conduction velocity (NCV) in the study population from the value in the individual and dividing the result by the SD. We constructed an ‘MNCV sum score’ based on MNCVs, an ‘SNCV sum score’ based on SNCVs, and a ‘total NCV sum score’ based on MNCVs, SNCVs and SNAPs. The sum scores combine information about the NCVs of different nerves by giving equal weight to each nerve, allowing for a more comprehensive assessment of nerve function in an individual [39, 40]. Primary analyses addressing a pre-planned hypothesis on NFL with primary and secondary exploratory endpoints were not adjusted for multiple testing.

In a secondary hypothesis-free exploratory analysis, all analyses mentioned above for NFL were performed for the remaining 60 biomarkers of the Olink Target 96 Neuro Exploratory multiplex assay. These hypothesis-free analyses were adjusted for multiple testing with the Bonferroni method (a recommended approach when many tests are carried out without pre-planned hypotheses) [41] and a Bonferroni-corrected p<0.0008 (0.05/61) indicated significant associations.

All statistical analyses were carried out with SAS version 9.4 (SAS Institute, Cary, NC, USA), and p values <0.05 were considered indicators of a statistically significant correlation, difference or association unless otherwise stated. The visualisation was carried out with RStudio version 4.0.5 (https://posit.co/download/rstudio-desktop).