European Child & Adolescent Psychiatry | 2021
Taming the chaos?! Using eXplainable Artificial Intelligence (XAI) to tackle the complexity in mental health research
Abstract
Mental disorders cause a significant degree of burden to affected individuals and to society at large. Reasons for this are their high prevalence (one in every two people suffers from a mental disorder at some point in their lifetime), their usually early onset (three in four patients fall ill before the age of 23), and-particularly if left untreated-their mostly chronic course, precipitating numerous disease-related disabilities and poor health outcomes [1–3]. In addition, a substantial percentage of non-responders and non-compliant patients exists. Notably, particularly for youth under the age of 18, access to diagnosis, prevention/intervention, and care for mental health problems is still relatively limited compared to common somatic issues [4]. In the following, we will explain the reason for this discrepancy and provide a possible solution. In general, the combination of clinical and translational science within medicine has steadily increased over the years, and this has led to tremendous progress also in mental healthcare [5, 6]. Nevertheless, mental disorders are very heterogeneous, dynamic, and multi-causal phenomena. Despite the widespread recognition of their inherent complex nature (including gene-environment and psyche-soma interactions as well as developmental and other experience-based changes over the life span), progress in understanding mental disorders as multifaceted bio-psycho-social conditions remains rather slow. In addition, even if acknowledged as such, our knowledge about etiopathophysiology, diagnosis, and management of mental disorders is still incomplete. The reasons that we still lack a more comprehensive picture of mental disorders are manifold, including the dichotomy of hypothesis-driven versus exploratory data-driven research methods and resulting findings, which all have their own pros and cons [7]. Furthermore, there is still an ongoing discussion to what extent the classification systems, such as DSM or ICD, are valid for diagnosing mental disorders [8]. For example, despite great research efforts clinically usable biomarkers that could potentially improve the early identification of mental disorders or that could be utilized for early intervention strategies (e.g., as predictors for treatment response) are still lacking. Among other reasons, the nosological classification systems primarily define mental disorders categorically according to a set of core symptoms, thereby neglecting the substantial dimensional, multifactorial, and heterogeneous clinical presentation and emergence of mental disorders as well as their symptomatic overlap. Consequently, dimensional transdiagnostic approaches have been introduced into the research arena, including the Research Domain Criteria (RDoC) project [9], but these approaches are still in their infancy, and they are no less hotly debated than the “traditional” ones [10]. This increase in complexity is accompanied by substantial technological and methodological advances in the areas of (epi)genetics, neuroimaging, psychophysiology, and others. However, these highly promising new leads that each attempts to identify the (neuro)biological correlates of mental disorders-or specific valid disease subtypes-have yet failed to convincingly resolve the issue of within-and acrossdisorder heterogeneity as they often lack specificity [8]. For instance, despite thorough research in the area of neuroimaging, with the exception of a few and relatively rare conditions, we do not have a single indicator available grounded in brain biology that can reliably distinguish patients with a specific mental disorder from typical controls, let alone * Veit Roessner [email protected]