Abstract

So far, we lack diagnostic tests for non-invasive identification of subject-specific pathophysiological pathways in psychiatric spectrum diseases. As a consequence, we are presently neither able to obtain mechanistically interpretable diagnoses for individual patients nor to make principled predictions about individualized treatment. Here, I propose a neurocomputational framework as a promising approach to address this critical problem for psychiatry. This framework – translational neuromodeling – combines physiologically interpretable dynamic system models with computational (Bayesian) models that are fitted to neuroimaging and behavioral data to provide estimates of pathophysiological mechanisms at the circuit and synaptic level. Subsequently, such model-based quantitative characterizations of “hidden” neuronal disease mechanisms can be exploited by machine learning techniques (e.g., Bayesian model selection and generative embedding) to generate probabilistic predictions about clinical outcome and treatment responses in individual patients. This presentation outlines the theoretical foundations of this framework and illustrates its potential by initial examples from clinical and pharmacological studies.