LAUSR

A decade ago, at a major international conference, we vividly remember a symposium on psychiatric Artificial Intelligence (AI) drawing a crowd of seven people— including the four speakers. Today one might get the impression that every other funding proposal is required to include at least some degree of AI-based analyses. At a time when 17 of the 29 hot topics listed by the most recent Gartner Hype Cycle for emerging technologies [1]—an indicator of perceived innovation—are either AI technologies (e.g., Generative Adversarial Networks) or include AI as a core component (e.g., autonomous driving), there is no shortage of promises. In both psychiatry and medicine in general, expectations to move beyond classical group-level statistics and enter the promising future of personalized medicine are high. Although AI has not yet fully hit mainstream psychiatric research, the availability and advancement of technology and methods have indeed led to a growing adoption of AI methods and agreement on best practice [2–4]. Despite this progress and some promising first applications (e.g., in suicide prediction [5]), translation to clinical practice has been hampered by a large degree of estimate variability and diagnostic heterogeneity. In the following, we will argue that the current drawbacks in psychiatry arise not primarily from a lack of methodological advancement and genuine clinical potential for AI in psychiatry, but from fundamental issues pertaining to sample size, model construction, evaluation practice, and the conceptualization of mental disorders. To overcome these challenges, we outline concrete steps towards an AI ecosystem that promotes a coordinated collaboration within the field to pave the way for better translation of AI solutions into clinical practice (see Fig. 1). Built upon four supporting pillars; collection, construction, evaluation, and translation, this ecosystem provides a framework for data collection, harmonization and sharing, guidelines for model construction, evaluation and distribution, methods for parsing heterogeneity as well as ethical and transparency standards.