Single-subject Prediction: A Statistical Paradigm for Precision Psychiatry

Abstract

Psychiatric patients are treated every day by medical doctors based on clinical guidelines grounded in group definitions. Therapeutic intervention for a particular patient, however, frequently follows a trial-and-error path (cf. Rush et al. 2006). On average, only about 50% of patients benefit from a specific psychotropic drug therapy (Wong et al. 2010). Similar failure rates apply to common psychotherapeutic treatments (Hofmann et al. 2012). To render clinical care more effective, brain-imaging and genomics are among the most promising, yet expensive avenues. In research on the neural and genetic basis of psychiatric disease, the prevailing research ideology aims to discover new pathophysiological mechanisms as a stepping-stone to then reduce suffering of psychiatric patients (Insel and Cuthbert 2015). Instead of trying to exploit newly discovered disease mechanisms toward novel treatments that help patient groups on average, an alternative research agenda is coming into reach over recent years. A fast and cost-effective strategy is to accurately predict which of the currently existing treatment options is likely to work best for one particular patient (Bzdok and Meyer-Lindenberg 2018; Perna and Nemeroff 2017; Stephan et al. 2017b). Increasing individualization of therapeutic intervention in precision psychiatry would also open up the opportunity to automatically derive the diagnosis or expected disease course in single patients from data. Because psychiatric disorders result from disturbed brain biology, quantitative measures from in-vivo brain-imaging is ready for such predictive modeling approaches because structural, functional, or diffusion magnetic resonance imaging (sMRI, fMRI, dMRI) and sometimes positron emission tomography (PET) are already available in most modern psychiatric hospitals. Moreover, such spatially or temporally highly resolved imaging techniques produce data of sufficient quantity and information granularity on which to apply data-driven statistical techniques (Eyre et al. 2016).

predictions, but not inferences, forecast what will happen.

White (1971)