Data Mining Framework for Discovering and Clustering Phenotypes of Atypical Diabetes

Abstract

Context: Some individuals present with forms of diabetes that are "atypical" (AD), which do not conform to typical features of either type 1 diabetes (T1D) or type 2 diabetes (T2D). These forms of AD display a range of phenotypic characteristics that likely reflect different endotypes based on unique etiologies or pathogenic processes.

Objective: To develop an analytical approach to identify and cluster phenotypes of AD.

Methods: We developed Discover Atypical Diabetes (DiscoverAD), a data mining framework, to identify and cluster phenotypes of AD. DiscoverAD was trained against characteristics of manually classified patients with AD among 278 adults with diabetes within the Cameron County Hispanic Cohort (CCHC) (Study A). We then tested DiscoverAD in a separate population of 758 multiethnic children with T1D within the Texas Children's Hospital Registry for New-Onset Type 1 Diabetes (TCHRNO-1) (Study B).

Results: We identified an AD frequency of 11.5% in the CCHC (Study A) and 5.3% in the pediatric TCHRNO-1 (Study B). Cluster analysis identified 4 distinct groups of AD in Study A: cluster 1, positive for the 65 kDa glutamate decarboxylase autoantibody (GAD65Ab), adult-onset, long disease duration, preserved beta-cell function, no insulin treatment; cluster 2, GAD65Ab negative, diagnosed at age ≤21 years; cluster 3, GAD65Ab negative, adult-onset, poor beta-cell function, lacking central obesity; cluster 4, diabetic ketoacidosis (DKA)-prone participants lacking a typical T1D phenotype. Applying DiscoverAD to the pediatric patients with T1D in Study B revealed 2 distinct groups of AD: cluster 1, autoantibody negative, poor beta-cell function, lower body mass index (BMI); cluster 2, autoantibody positive, higher BMI, higher incidence of DKA.

Conclusion: DiscoverAD can be adapted to different datasets to identify and define phenotypes of participants with AD based on available clinical variables.

Keywords: atypical diabetes; bioinformatics; clusters; ketosis-prone diabetes; type 1 diabetes; type 2 diabetes.

Figure 1.

Identification of atypical diabetes (AD) participants for the Cameron County Hispanic Cohort (CCHC) (Study A) using DiscoverAD. (A) Flowchart of CCHC population. (B) The first filter step (exclusion for T1D and T2D). (C) The second filter step (inclusion for AD).

Figure 2.

Flow chart for identification of atypical diabetes (AD) participants for the Texas Children's Hospital Registry for New-Onset Type 1 Diabetes (TCHRNO-1) (Study B) using DiscoverAD. (A) The first filter step (exclusion for T1D and T2D). Participants with missing data for filtering variables (n = 7 missing disease duration, n = 9 missing BMI) were identified as atypical for subsequent manual review. (B) The second filter step (inclusion for AD).

Figure 3.

(A) DiscoverAD identified clusters’ radar plots for the Cameron County Hispanic Cohort (CCHC) (Study A). Continuous variables presented as median of scaled values using min-max normalization (age at onset [years], age at visit [years], body mass index (BMI) [kg/m²], C-peptide levels [ng/mL], GAD65Ab levels [U/mL], hip circumference (cm), Homeostasis Model Assessment – Beta-cell Function [HOMA-beta], Homeostasis Model Assessment – Insulin Resistance Index [HOMA-IR], insulin [mU/L], mean fasting blood glucose [MFBG] [mg/dL], waist circumference [cm]). Dichotomous variables presented as mean (diabetic ketoacidosis [DKA] [0, no DKA history; 1, past DKA history], GAD65Ab-positive, gender [0, female; 1, male]). (B) DiscoverAD identified clusters’ radar plots for the Texas Children's Hospital Registry for New-Onset Type 1 Diabetes (TCHRNO-1) (Study B). Continuous variables presented as median of scaled values using min-max normalization (bicarbonate [mmol/L] at diagnosis, body mass index [BMI] percentile (%), beta-hydroxybutyrate [BOHB] (mmol/L) at diagnosis, age at diagnosis [years], C-peptide levels [ng/mL] at diagnosis, the 65 kDa glutamic acid decarboxylase [GAD65] autoantibody levels [U/mL] at diagnosis, glucose [mg/dL] at diagnosis, insulin autoantibody [IAA] levels [U/mL] at diagnosis, ICA512 autoantibody levels [U/mL] at diagnosis, pH at diagnosis). Dichotomous variables presented as mean (autoantibody [GAD65, IAAs, or ICA512] positivity, diabetic ketoacidosis [DKA] [0, no DKA history; 1, past DKA history], gender [0, female; 1, male]).