Introduction: Insulin omission is a specific eating disorder that occurs in patients with type 1 diabetes mellitus (T1DM), mostly in females, who omit or restrict their required insulin doses in order to lose weight. Diagnosis of this specific eating disorder is difficult. We aimed to use multiple clinical criteria to create an algorithm for the detection of insulin omission eating disorder.
Methods: Data from 287 (181 females and 165 males) patients with T1DM and 26 patients with T1DM and eating disorders were used. The Weka (Waikato Environment for Knowledge Analysis) machine learning software, decision tree classifier based on the J48 algorithm with 10 fold cross validation was used to developed prediction models. Model performance was assessed by cross-validation in a further 43 patients.
Results: The insulin omission eating disorder and non-eating disorder populations were discriminated by: female sex, HbA1c>9.2%, more than 20% of HbA1c measurement above the 90th percentile, the mean of 3 highest delta HbA1c z-score >1.28, current age and age at diagnosis. The models developed showed good discrimination (sensitivity and specificity 0.88 and 0.74, respectively). The external test dataset revealed good performance of the model with a sensitivity and specificity of 1.00 and 0.97, respectively.
Conclusions: using data mining methods we developed a clinical prediction model to determine an individual’s probability of having eating disorders. This model provides a decision support system for the detection of insulin omission eating disorders in adolescent females with T1DM.
