Measurement of Heterogeneity in Computational Psychiatry

Abstract

We introduce representational Rényi heterogeneity (RRH), which generalizes existing heterogeneity measurement approaches from ecology (biodiversity measures) and economics (inequality measures). We show that RRH retains the interpretability of the standard family of Rényi heterogeneity indices, while enabling heterogeneity measurement in datasets of nearly arbitrary form.

In the applied section of this thesis, we present the largest machine learning (ML) based study of prediction of mood stabilizer treatment in bipolar disorder (BD) based on clinical features. Using our RRH framework, we derived a method called exemplar scoring, which enabled us to identify “canonical” clinical profiles of lithium responsive and non-responsive BD, respectively. We then show that lithium response is more easily predicted genetically among individuals with canonical clinical profiles.

An ancillary contribution of this thesis is demonstration of a scenario in which dichotomization of a continuous variable yields more information than retaining the continuous representation.