Beyond completion time: predicting individual differences in executive functions using cursor trajectory features in an online Trail Making Test

BackgroundThe Trail Making Test (TMT) is a widely used instrument for assessing executive functions due to its sensitivity. But, its traditional scoring, based solely on total completion time, limits its specificity by losing the rich behaviour required to complete the task, involving integrating visual search, motor planning and task switching. A computerised implementation of the TMT (cTMT) allows high-resolution cursor trajectories to be recorded, offering access to fine-grained features of the trajectory, while an online administration improves accessibility and statistical power. ObjectiveThis study aimed to extract high-resolution cursor trajectory features from an online cTMT and evaluate their capacity to predict individual differences in core executive functions, such as visual working memory (VWM), inhibitory control and age, using machine learning, as well as validating the feasibility of a fully online data acquisition pipeline as a basis for digital biomarkers. MethodsParticipants completed an online battery comprising the cTMT and three validation tasks: the Change Detection Task (CDT), the Stop-Signal Task (SST) and the Go/No-Go task (GNG). A total of 104 features (26 features x Part A/B x whole-trial/first-10-targets) were extracted from cursor trajectories, including trajectory profiles and segmentation into latent states: Search, Travel and Hesitation. Regression models were trained within a nested Leave-One-Out cross-validation framework with inner 10-fold hyperparameter tuning, feature selection and standardisation applied strictly within folds. Performance was assessed via mean absolute error (MAE), normalised error (MAE/SD) and permutation testing; SHapley Additive exPlanations (SHAP) were used to characterise feature importance. ResultscTMT features strongly predicted age across all models (p < .001), with MAEs roughly one-third lower than the targets dispersion, driven predominantly by distance-based "circuitousness" metrics from both parts. GNG accuracy and c-coefficient were both significantly predicted, but with a dissociated pattern: accuracy was best explained by Part B (alternation) metrics, whereas the c-coefficient was almost exclusively predicted by Part A (simple sequencing) metrics. SST response time (SSRT) was not significantly predicted by any model. VWM, characterised by the mean Cowan s K, was significantly predicted mainly by the more complex models, and involving state transitions and search phases in both Part A and B. ConclusionsMoving beyond completion time, cursor trajectory dynamics from an online cTMT provide a rich behavioural signal for predicting age, VWM capacity and distinct aspects of inhibitory control. The dissociation between Part A and Part B predictors supports differentiated cognitive processes within the TMT and positions the cTMT as a scalable, portable digital biomarker with promise for computational psychiatry and personalised neuropsychology.