Predictive Analytics from a Mental Health Perspective

While there is a plethora of biomarkers for diagnosis and treatment selection in somatic medicine, no comparable biological or psychological markers are available in mental health. Following the bio-psycho-social model, both the pathogenesis and treatment effects are determined by many concurring factors. With machine learning, predictive analytics offer a promising set of tools for translating patterns and interactions in and between a variety of variables into a conclusion for the individual patient. These methods "learn" the association between predictors and outcomes from already available data and can then apply the resulting model on new data, for which the outcome is still open. However, it is crucial to evaluate beforehand whether the learned model is meaningful. To illustrate this approach, we present a number of studies that used predictive analytics for diagnostics, for predicting risk trajectories and for predicting psychotherapy treatment outcomes ("theranostics"). Their results are promising, but prior to clinical practice the prediction accuracy has to be increased and tested in different settings and populations. For increasing prediction performance, combining several data modalities, such as clinical, neurostructural, functional and genetic data, and focusing on variables that map mechanisms of psychopathology and change, are reasonable. Moreover, teaming up with clinician and patient representatives is recommended for increasing the acceptance of such markers and discussing the ethical and societal implications of predictive analytics in mental health. If successful, predictive analytics bear the potential to increase diagnostic reliability particularly in challenging cases, to identify potentially negative trajectories early on and to support allocating patients to their individually optimal treatment. (c) 2020 S. Karger AG, Basel