Differentiating mania/hypomania from happiness using a machine learning analytic approach.

Background: This study aimed to improve the accuracy of bipolar disorder diagnoses by identifying symptoms that help to distinguish mania/hypomania in bipolar disorders from general 'happiness' in those with unipolar depression. Methods: An international sample of 165 bipolar and 29 unipolar depression patients (as diagnosed by their clinician) were recruited. All participants were required to rate a set of 96 symptoms with regards to whether they typified their experiences of manic/hypomanic states (for bipolar patients) or when they were 'happy' (unipolar patients). A machine learning paradigm (prediction rule ensembles; PREs) was used to derive rule ensembles that identified which of the 94 non-psychotic symptoms and their combinations best predicted clinically-allocated diagnoses. Results: The PREs were highly accurate at predicting clinician bipolar and unipolar diagnoses (92% and 91% respectively). A total of 20 items were identified from the analyses, which were all highly discriminating across the two conditions. When compared to a classificatory approach insensitive to the weightings of the items, the ensembles were of comparable accuracy in their discriminatory capacity despite the unbalanced sample. This illustrates the potential for PREs to supersede traditional classificatory approaches. Limitations: There were considerably less unipolar than bipolar patients in the sample, which limited the overall accuracy of the PREs. Conclusions: The consideration of symptoms outlined in this study should assist clinicians in distinguishing between bipolar and unipolar disorders. Future research will seek to further refine and validate these symptoms in a larger and more balanced sample.

Automated summary

The study used machine learning to identify highly discriminating symptoms between bipolar disorder and unipolar depression patients, assisting clinicians in distinguishing between the two conditions. Despite the unbalanced sample, the prediction rule ensembles showed potential in accuracy and may supersede traditional classificatory approaches.