Can 7-Tesla structural MRI morphology predict Trail-Making Test performance?

A median split, not a stopwatch prediction

The paper's headline accuracy of 0.795 does not mean the model estimated how many seconds a person would take on the Trail-Making Test. The Trail-Making Test asks participants to connect items in order; part A uses numbers and part B adds alternating letters and numbers, making part B the more cognitively demanding task. The authors combined TMT-A and TMT-B measurements and trained a deep learning model to distinguish participants whose completion times fell above the median from those below it. The outcome is therefore a two-group classification within this sample, not a continuous time prediction and not a diagnostic label.

That distinction matters for reading every later number. An accuracy of 0.795 describes how often the model assigned the correct side of the median split, and the paper reports it as very good. It says nothing about the size of the time difference between the two groups, and it does not establish that the same threshold would separate fast from slow performers in another sample.

74 patients and 41 controls at 7 Tesla

The study measured 74 transdiagnostic patients and 41 healthy controls with a 7-Tesla MRI system. "Transdiagnostic" here means the patient group was not restricted to a single psychiatric diagnosis; the authors argue that symptom clusters and underlying processes cut across diagnostic categories, so a sample defined by diagnosis would lose information. The comparison group of 41 healthy controls provides the reference against which patient morphology and test performance were examined.

Morphology was summarized with three measures computed for parcels of the Schaefer 100 Atlas: grey matter volume, cortical thickness, and gyrification, which describes the degree of cortical folding. These parcel-level values, not raw images, were the inputs to the deep learning model. The model's job was to map that morphological profile onto the above- versus below-median TMT class.

Gyrification carried the most predictive weight

A SHAP analysis was used to identify which input variables drove the model's predictions. SHAP values attribute each prediction to individual features, so a large contribution means that feature moved the model's output substantially for that case. The most decisive feature was gyrification. Grey matter volume in temporal and parietal-occipital regions also contributed, while cortical thickness did not appear among the most important features at all.

The authors interpret this ranking through the different dynamics of the three measures. In adults, grey matter volume and cortical thickness change over time to a much greater degree than gyrification, so the authors suggest volume and thickness are better suited to capturing temporary changes, whereas folding patterns are more likely to reflect predisposing structure laid down early in development. That is the authors' explanation of the feature ranking, not a demonstrated causal pathway. The paper also notes that variance was fairly evenly distributed across variables, with no single brain area showing dominant predictive power.

Frontal, temporal and parietal regions marked group differences

The areas that indicated inter-group differences were concentrated in the frontal cortex, along with temporal and parietal regions. Through the SHAP analysis the authors located contributions in the somatomotor network, the limbic network, the dorsal attention network, and the default mode network. The gyrification and grey matter volume of the dorsal prefrontal cortex, the gyrification of the lateral prefrontal cortex, and the gyrification of the left orbitofrontal cortex were among the features that contributed to classification accuracy.

The paper connects these regions to task demands rather than treating them as isolated markers. Completing the TMT requires motor coordination and fine control of speed, which makes involvement of the somatomotor system intuitive to the authors. The orbitofrontal cortex is linked in the cited literature to temporal cognition and sequencing, abilities the test draws on. The authors describe the overall pattern as a complex set of morphological alterations overlapping with areas changed by pathology, and they frame the brain as a functional system in which problems in one area can accompany reduced performance elsewhere.

Diagnosis was a poor guide to TMT performance

A sizable proportion of both the healthy participants and the patients had above-median TMT-A and TMT-B times. In other words, falling on the slow side of the split was not confined to the patient group, and the authors conclude that diagnosis was not a reliable predictor of low cognitive function as measured by the TMTs in this sample.

This observation is the empirical basis for the paper's transdiagnostic argument. If diagnostic category poorly separates people by test performance, then a measurement-based grouping, here morphological features plus test results, may capture variation that diagnosis obscures. The authors present this as support for a transdiagnostic approach in clinical practice and for dimensional descriptions of pathology, and they position the model as a first step toward identifying variables that could predict which biological interventions are likely to improve cognitive function.

A classification result, not a clinical tool

Several features of the design bound what the 0.795 accuracy can support. The target is a median split computed within this mixed sample of 74 patients and 41 controls, so the model is separating people relative to this group's midpoint rather than against an external clinical cutoff. The sample is modest and transdiagnostic, which is central to the authors' argument but also means the result describes this heterogeneous group rather than any single diagnostic population.

The authors report relatively low standard deviations across measurements after repeated hold-out loops and treat this as supporting generalizability, and they describe the model's reliance on regions already implicated in the literature as validating the data-driven approach. Those are the authors' assessments of reliability. The paper does not report testing the model on an independent cohort, so the accuracy figure should be read as internal performance on this sample rather than as an expected accuracy in a new clinical setting.

Folding patterns point to earlier, more stable structure

The most consequential result for readers is not the accuracy figure alone but the split among the three morphological measures. If gyrification dominates while cortical thickness contributes nothing to the top features, the model is leaning on structural patterning that the authors associate with ontogenesis and early childhood development, with grey matter volume in temporal and parietal-occipital regions adding a more modifiable component. That reading implies the predictive signal is not simply a snapshot of current brain state.

The authors use this to argue that grey matter volume and cortical thickness may be more useful for tracking temporary change, while gyrification may index longer-standing predisposition. They illustrate the point by suggesting that a long-term reduction in cognitive capacity from chronic stress stretching back years would not strongly influence the degree of cortical folding. This remains an interpretation rather than a tested claim in the study, but it is the specific priority the paper leaves readers with: when comparing morphological predictors of cognition, separate measures with different temporal dynamics instead of treating brain structure as one variable.

Sources used in this answer