LAUSR

Scant attention has been paid to evaluating differences in the prevalence of comorbidities and diabetes-related complications in familial versus sporadic type 1 diabetes (1). Knowledge gains in this area could advance the development of risk prediction tools and tailored interventions for preventing or delaying onset of comorbidities or diabetes-related complications in high-risk patient subgroups. To address this gap, we applied a computationally optimized, exploratory data mining algorithm to the T1D Exchange Clinic Registry (2). For the first time in a large U.S.-based cohort, we assessed demographic and phenotypic factors and comorbid conditions for associations with familial (i.e., having an affected first-degree relative) or sporadic (i.e., having no family history of type 1 diabetes) disease. The T1D Exchange Clinic Registry is a deidentified, publicly available data set comprising 34,013 adult and pediatric participants who received routine clinical care at 83 U.S.-based endocrinology practices between July 2007 and April 2018 (3). We analyzed participants with a family history of type 1 diabetes involving a first-degree relative, i.e., father (n = 1,464), mother (n = 818), sibling/twin (n = 1,882), and/or child (n = 228) (total n = 3,941) or no family history of type 1 diabetes (n = 12,291). Excluding participants >50 years old resulted in a relatively balanced distribution of age and diabetes duration across both subgroups. A contrast pattern mining algorithm detects significant differences in the frequencies of attributes across two patient subgroups. We used our validated algorithm to discover individual and co-occurring characteristics that were documented significantly more frequently in familial versus sporadic type 1 diabetes. Here, we refer to these characteristics as “patterns” or “feature patterns.” Our algorithm returns feature patterns consisting of one, two, or three elements. Individual elements are synonymous with individual characteristics. Metrics used in feature pattern analysis include support, growth, and confidence (4,5). Support is the proportion of individuals in a subgroup who are associated with a given feature pattern. Growth is a support ratio between subgroups. Confidence corresponds to the statistical concept of positive predictive value. We used Fisher exact tests to calculate the statistical significance of each pattern (P < 0.05) and the Benjamini-Hochberg (BH) procedure to control for false discovery (false discovery rate of 0.1). Of 16,232 individuals who met inclusion criteria, 24.3% (n = 3,941) had an affected first-degree relative. Median age of familial cases was 18 (interquartile range [IQR] 15, 27) years; for sporadic cases, median age was 18 (IQR 15, 23) years (P = 0.05). Median diabetes duration in familial cases was 10 (IQR 6, 16) years; in sporadic cases, median diabetes duration was 9 (IQR, 6, 14) years (P < 0.001). Median age at diagnosis was 8 (IQR 4, 12) years in both subgroups (P = 0.002). Mean (± SD) hemoglobin A1c (HbA1c) for familial cases was 8.4 ± 1.3% (68.7 ± 14.7 mmol/mol); for sporadic cases, mean HbA1c was 8.3 ± 1.2% (66.72 ± 13.2 mmol/mol) (P < 0.001). We discovered 590 feature patterns that met a minimum prevalence threshold of 1% in at least one subgroup. After controlling for false discovery, 265 patterns retained statistical significance. These included 29 single-element patterns, 103 two-element patterns, and 133 three-element patterns (Table 1). Conditions that were significantly enriched in familial type 1 diabetes included hypertension, hyperlipidemia/