Computational factor modelling framework and behavioural task.

Computational factor modelling approach: a. Abstract representation learning task: Participants learned the fruit preferences of Pac-Man-like characters, with the features that predict preference changing block-by-block. Three types of feature combinations could predict preference: colour–stripe orientation, colour–mouth direction, and stripe orientation–mouth direction (see the example in the top-right box). The cursor on the slider allows participants to report both their choice and their confidence simultaneously. Latent behavioural scores are estimated through computational modelling. b. Transdiagnostic psychiatric dimension scores are extracted across diagnostic boundaries constructed by factor analysis. Then, extracted latent behavioural variables can be linked to each transdiagnostic psychiatric dimension. This data-driven approach enables an integrated understanding of how computational mechanisms underlying key behaviour attributes map onto transdiagnostic symptom dimensions.

Behaviour validation of the abstract representation learning task.

a. Trial-by-trial ratio of correct responses served as an index of within-block learning. Each dot represents the mean across participants, error bars denote the standard error of the mean (SEM), and the shaded area indicates the 95% confidence interval. Ratio-correct values were computed for each trial between completed blocks and then averaged across participants. b. Relationship between learning speed and time across participants. Learning speed was defined as the inverse of the maximum-normalised number of trials required to complete a block. Thin grey lines show individual participants’ data, and the bold black line represents the group-average fit. Circles represent population-level mean values, and error bars represent the SEM. c. Averaged confidence ratings were positively associated with learning speed across participants. Each point corresponds to a single participant, and the solid line represents the regression fit. ** p < 0.01.

Correlations between item loadings for the three-factor structure from de-novo factor analysis between a collected sample in this study (N = 249) and a larger reference dataset (N = 19,505).

Scatter plot of weights of each transdiagnostic factor between the newly collected global sample (from this study, x-axis) and the reference sample (from a large online survey in Japan(Oka et al., 2025), y-axis). Each circle indicates a single questionnaire item. Correlation values are as follows: FA1 r = 0.63, 95% CI [0.53, 0.71], p < 0.001. FA2 r = 0.44, 95% CI [0.31, 0.55], p < 0.001, FA3: r = 0.46, 95% CI [0.34, 0.57], p < 0.001.

Behaviour modelling framework.

a. Feature RL model. The model has 23 = 8 states, each corresponding to a combination of all three features. b. Abstract RL model. The model is designed with 22 = 4 states, where each state is defined by the combination of two features, with three possible ways to achieve this 4-state configuration. c. Histogram of probability of using the abstract strategy across participants. The vertical line shows the population mean (M = 0.26).

Associations between task-related behaviour measures and transdiagnostic symptom dimensions.

Regression coefficients between each metric and dimension scores based on bootstrap multiple linear regression estimation with 5000 iterations. a. Associations between the abstraction parameter and each dimension score. b. Associations between the averaged confidence and each dimension score. c. Associations between the metacognitive sensitivity and each dimension score. The black vertical line in each histogram represents the hypothetical null effect (coefficient = 0). The red vertical dashed line in each histogram represents the average beta coefficient. ** pboot < 0.01