Machine Learning Reveals Neural Markers of Motivation

A doctoral thesis by Nam Trinh applies machine learning to multimodal neuroimaging to probe motivated behaviour, from ADHD to individual differences in effort and reward sensitivity. The work reports three complementary studies using EEG, diffusion MRI, and T1-weighted MRI that together identify neural signatures tied to clinical and subclinical motivational phenotypes. Key convergent loci are fronto-parietal circuits and SMA-connected white matter pathways.

Study 1 trained classifiers on task-based and resting EEG recorded during a stop signal task; models trained on task data outperformed resting-state models, with predictive signal concentrated in gamma-band spectral power over fronto-central and parietal electrodes. Study 2 used whole-brain permutation-based analyses on diffusion MRI to link white matter integrity to computational model parameters for effort and reward sensitivity, highlighting SMA-connected tracts as hubs. Study 3 analysed grey matter volumes from structural T1-weighted MRI to decode reward sensitivity and subclinical apathy, with machine learning confirming robust decoding for reward-related traits.

Why it matters The work provides multimodal candidate biomarkers that bridge computational constructs (effort, reward sensitivity) and measurable brain features. For practitioners, the paper demonstrates that task-evoked EEG gamma features and specific white matter pathways carry actionable signal for classification and dimensional prediction. The approach aligns computational modelling of behaviour with imaging-based decoding, a practical blueprint for translational neuroimaging.

Implications and limitations The findings are promising for diagnostic augmentation and personalised interventions, but the abstract does not report sample sizes, cross-site replication, or out-of-sample generalisation metrics needed to judge clinical readiness. Effect sizes, model architectures, feature selection pipelines, and robustness to confounds will determine real-world utility.

Follow-up work for replication, open-source code and trained models, and extensions that combine modalities in unified predictive frameworks. External validation on larger, multi-site cohorts will be the decisive next step for these biomarkers.