Deep Learning of dynamical systems for mechanistic insight and prediction in psychiatry

Daniel Durstewitz
Ruprecht-Karls-Universität Heidelberg

Dynamical systems theory (DST) provides a unifying formal language within which many natural and real world phenomena at multiple time scales can be described, analyzed, and predicted. In particular, neural and behavioral phenomena evolve in time, and this temporal dynamics is essential for understanding the underlying processes and computations, and for predicting future states of the system. Many observations in neuroscience and psychiatry naturally come as multivariate time series, like fMRI, EEG or MEG recordings, or like behavioral measurements taken across trials in experiments, or as recorded through wearable sensors or ecological momentary assessments (EMA) via mobile devices. If we had direct access to the underlying dynamical systems generating these time series, we could systematically analyze the governing mechanisms, obtain functional insights, and predict how the system would continue to evolve in the future.
In recent years, we have developed mathematical tools based on deep generative recurrent neural networks (RNN) for inferring dynamical systems directly from empirical data. These AI-based tools enable to construct powerful generative models of individual brains or behaviors from which we can obtain novel diagnostic markers, rooted in dynamical properties, can forecast individual disease trajectories, and can simulate the effect of pharmaceutical or other therapeutic interventions for guiding personalized treatment. Example applications on fMRI and EMA data will be presented.

Presentation (PDF File)

Back to Computational Psychiatry