Virtual reality (bio-)markers differentiate ADHD from depression and controls.

The heterogeneity of attention-deficit/hyperactivity disorder (ADHD) and its symptom overlap with other mental disorders limit the clinical utility of isolated diagnostic markers. Virtual reality (VR)-based paradigms offer potential to aid individual diagnostic decisions. This study investigated the diagnostic value of classification models based on multimodal data derived from a virtual email sorting task (VEST). Data were pooled from two studies and used to distinguish unmedicated participants with ADHD (n = 43) from medicated participants with ADHD (n = 23), participants with major depressive disorder (MDD, n = 16), and control participants (n = 21). During the VEST, objective measures of task performance, head actigraphy, and eye tracking were recorded. Subjective data were obtained by assessing task-related ADHD symptoms. Using nested cross-validation, support vector machines were trained and tested, comparing model performance between objective, subjective and combined parameters. The highest balanced accuracy for the classification of participants with ADHD versus control participants (0.85) was achieved by a single-feature model using self-reported inattentiveness. Participants with ADHD and MDD were differentiated with 0.75 balanced accuracy using a 13-feature model combining subjective and objective measures. A 13-feature model derived exclusively from objective measures yielded 0.64 balanced accuracy for the differentiation of medicated versus unmedicated participants with ADHD yet did not perform significantly above chance level. These findings underscore the potential of VR-based multimodal assessments and machine learning to advance precision diagnostics and highlight the utility of self-report measures to detect ADHD while demonstrating the diagnostic potential of objective measures to distinguish disorders and capture medication effects.