Melancholia defined with the precision of a machine

Abstract

BackgroundThe status of melancholia as a categorical or dimensional condition remains unclear, and no measure of melancholia has achieved definitive status. This study aimed to use a machine learning approach to assess whether a pre-established cut-off score on the Sydney Melancholia Prototype Index (SMPI) provided clear differentiation of melancholic/non-melancholic depression, and to identify the items making the most distinct contribution. MethodsWe analysed amalgamated data sets of 1513 clinically depressed patients assessed via the clinician-rated version of the SMPI (SMPI-CR). We also evaluated the self-report version of the SMPI (SMPI-SR) in a combined clinical/community sample of 2025 depressed patients and senior high school students. Rule ensembles were derived in which the outcome measure was the presence/absence of melancholia (defined as scoring at or above a SMPI cut-off score that had been established in previous studies) and the predictive variables were the individual SMPI items. ResultsThe pre-established SMPI cut-off score was confirmed as differentiating melancholic/non-melancholic with near perfect accuracy for the SMPI-CR, and with very high accuracy for the SMPI-SR. The relative importance of all SMPI items was quantified. LimitationsIt is difficult to validate SMPI-assigned diagnoses due to the lack of any similar measures. ConclusionsThe SMPI-CR was confirmed to be a highly precise instrument for differentiating melancholic and non-melancholic depression. Its use will advance clinical decision making and studies evaluating causes, mechanisms and treatments for the two depressive sub-types, as well as assist clarification as to whether melancholia is categorically or dimensionally distinct from non-melancholic depression.