Beyond accuracy: interpretable behavioural profiling of children with ADHD using a visuospatial working memory game

Abstract

Attention-Deficit/Hyperactivity Disorder (ADHD) is usually diagnosed using rating scales and clinical interviews, which offer limited access to underlying cognitive mechanisms. Existing objective tasks often rely on coarse summary statistics or less interpretable feature spaces, making it unclear which behavioural properties distinguish ADHD from typical development. This study aims to identify interpretable behavioural features from a visuospatial working memory game. We collected click-level data from 174 children aged 7-10 years playing Honey Memory, a visuospatial working memory game. From each session, we engineered behavioural features encompassing different aspects of performance: working memory capacity measures, error-monitoring indices, behavioural variability metrics, response-time dynamics, and spatial search patterns. ADHDrelated differences spanned multiple cognitive domains. Children with ADHD showed reduced working memory capacity (shorter spans, lower accuracy), impaired error monitoring (more frequent and longer runs of consecutive errors), elevated behavioural variability (increased correct-incorrect response toggling, greater trial-to-trial fluctuations), and inefficient spatial search strategies (shorter, proximity-bound paths). These differences were most pronounced at medium difficulty levels. However, response time dynamics in general did not show a statistically significant difference between the two groups. Random Forest classifiers achieved high within-site accuracy (Best accuracy: 90.1%) and moderate cross-site generalisation. Together, these findings identify a multi-aspect behavioural profile of ADHD, with classification performance emerging as a downstream benefit of mechanistically interpretable features.