Multimodal fusion of gameplay telemetry and linguistic behaviour for three-tier ADHD risk stratification in adult gamers: a machine-learning study with cross-domain transfer evaluation

The deposit contains participant-level clinical profiles with ASRS-v1.1 scores and the three-level risk label (1,200 x 14), per-session behavioural telemetry from ranked MOBA matches (9,273 x 21) and from battle-royale matches (8,237 x 21), and the participant-level fusion matrix used for all reported modelling (1,200 x 53), of whose 43 candidate features 27 enter the models (13 telemetry, 11 text, 3 demographic; ASRS scores are excluded because they define the label). Three Python scripts reproduce every reported quantity in roughly eight minutes, covering model comparison across six classifiers under stratified five-fold cross-validation, bootstrap and nested cross-validation intervals, DeLong comparisons with Holm adjustment, a label-permutation null, variance inflation and penalised-regression robustness checks, source-composition probes and modality ablation; the 18 included reference tables are the verbatim output of this code run against this data. See DATA_README.md for the column dictionary, the fixed perturbation seeds required for exact reproduction, and run instructions.