• Aims: Identify potential clusters/groupings within neuropathy assessment items and responses that may characterize neuropathy in those with type 2 diabetes, using data-driven methods.
• Study sample: 10-year exam of ADDITION-DK.
• Assessment: Diabetic neuropathy assessed by several clinical scoring systems (TCSS, mTCSS, UENS, MNSI, DN4), Sural nerve conduction studies, and heart rate variability measurements.
• Analysis: Cross-sectional. Hierarchical clustering to find clusters between individuals’ responses. Factor analysis of mixed data to find groupings in assessment items. Used data on complete cases of all nerve assessments (n=183).
• Results: For 3 fixed clusters (Fig. A): 1 was dominated by “no”, “pass”, or “decreased” responses; 2 was dominated by responses for reduced sensation and greater pain in feet; 3 was dominated by responses for reduced touch, temperature, and vibration sensation and more pain up to the ankle. Factor analysis of first 3 components (11.1%, 6.5%, 5.3% mean explained variances, respectively) showed top contributor assessment items grouped into touch sensation, feeling of pain in feet, and ankle reflexes (Fig. B).

Additional details on methods

• Study design: We performed a cross-sectional analysis on those from the 10-year clinical examination of ADDITION-Denmark after a screening-based diagnosis of type 2 diabetes.
• Sequence of steps in analysis:
  • Data was split into two sets, a training set (85%) and a testing set (15%) for later cross-validation (CV) of the findings.
  • Hierarchical cluster analysis (HCA) was used cluster of individuals based on their responses to the neuropathy assessments.
  • Factor analysis of mixed data (FAMD) was used to determine the contribution of each assessment item to the underlying variance in the data.
  • The training data set was split into 10-fold CV sets, and this was repeated 20 times, for a total of 200 randomly sampled sets of data.
  • HCA and FAMD analyses were applied separately to each individual sampled set.
  • For each HCA computation, the individual’s cluster number was extracted. For each FAMD computation, the assessment item’s contribution to the components (as well as the explained variance) was extracted.
  • Results from all 200 analyses were combined together, providing for a distribution of possible values of the findings.
  • From the results, the most frequent/common assessment responses (for each HCA cluster) and items (for each FAMD component) were extracted.
  • Full analytic steps can be found in the code of the project repository, specifically in the distance-hclust.R, famd.R, and plan.R files.