A new machine learning-derived screening measure for differentiating bipolar from unipolar mood disorders

Background: While there are several accepted screening measures for identifying those with a bipolar disorder, variations in overall classification rates argue for the pursuit of a more discriminating measure. Extant measures, as well as the DSM-5, rate each diagnostic criterion as having equivalent weighting values; an approach which may compromise diagnostic assignment if symptoms vary considerably in their diagnostic sensitivity. We therefore sought to develop a new measure and examine whether a weighted rating scale was superior to one assigning equivalent weightings to each item. Methods: An international sample of 165 bipolar patients and a comparison sample of 29 unipolar patients completed a measure assessing 96 putative manic/hypomanic symptoms. A previous machine learning analysis had identified the twenty most discriminating items. In this study, analysis was undertaken involving only the ten most discriminating items. Results: Whether items were scored as each having equivalent value or as weighted by their machine learning-generated values, classificatory accuracy was extremely high (in the order of 96%). Analyses also identified optimal cut-off scores. High classificatory accuracy was also obtained when scores for separate bipolar I and bipolar II groups were compared with scores from the unipolar group. Limitations: The sample consisted of comparatively few unipolar patients. Conclusions: The ten-item set allows a new measure for researchers to evaluate, while the items should assist clinician assessment as to whether a patient has a bipolar or unipolar mood disorder.