A machine learning analysis of psychopathological features of Eating Disorders: a retrospective study

Background: The clinical presentation of Eating Disorders (EDs) is often characterized by a great phenotypic variability and by a substantial instability over time of diagnostic categories. For these reasons, it has been proposed a different approach to EDs, encompassing both their dimensional and categorical descriptions, in a lifetime perspective, namely the ‘Anorexic-Bulimic Spectrum’ (ABS). Here we report a retrospective study with the interview built and validated for the assessment of ABS signs and symptoms, the ‘Structured Clinical Interview for Anorexic-Bulimic Spectrum’ (SCI-ABS), administered together with the ‘Mood Spectrum Self Report’ (MOODS-SR), a questionnaire able to assess sub-threshold mood spectrum dysregulations often comorbid with EDs signs and symptoms. The main aim of the study was twofold: to assess and better characterize clinical phenotypes of EDs; to highlight potential lifetime sub-threshold mood dysregulations that might occur comorbid with EDs, and that patients might consider relevant to their ‘subjective experience of illness’. In order to obtain these goals, we decided to utilize a machine learning analysis. Methods: two groups were recruited and compared, namely patients with EDs (n=53) and healthy controls (HC) (n=54). Both groups underwent psychological testing with MOODS-SR and SCI-ABS. Results: in discriminating and classifying EDs individuals from HC, machine learning classifiers obtained an accuracy higher than 70%. Based on all variables considered, the analysis revealed that SCI-ABS ‘Phobias’ domain (more in detail, ‘Weight Gain Phobia’ total score), the ‘Impairment and Insight’ item 5, (‘…your relationship with food was all you could think about?’) and the MOODS-SR item 154 (‘you were less sexually active than is typical for you?’) were the best psychological elements in discriminating EDs patients from HC (accuracy range: 72.90-86.92%). Given the large number of predictors, we run a supervised attributes selection procedure. The procedure yielded an accuracy of 90.65% in classifying EDs patients from HC. Conclusions: the very high overall accuracy is indicative that the selected combinations of features capture the most important determinants in the discrimination of EDs patient’s vs HC. The items selected by the machine learning analysis confirmed that an extreme polarization of ideas on weight and food control characterize the cognitive asset of EDs patients.