Summary
People with and without epilepsy can have AI models that predict their brain activity equally well, but the internal ways these models understand the brain’s rhythms can be quite different. The authors used EEG data to train personalized models that look at brain activity patterns over time. They found that epilepsy patients’ brain models have more complex connections internally, even though the models’ accuracy is about the same as for non-epilepsy patients. This means a model’s predictions alone don’t tell the whole story—it’s important to study what the model learns inside, especially for personalized medical uses.
What this means in practice
- •For clinical data scientists: Identify differences in brain disorder models’ internal dynamics to improve personalized diagnosis and treatment strategies.
- •For medical device developers: Design EEG analytics tools that assess model-internal structures in addition to prediction accuracy for better clinical decision support.
Abstract
As AI models move toward clinical decision-making and personalized treatment, understanding \emph{what} a model learns is important beyond predictive accuracy alone. We investigate whether personalized latent dynamics reveal clinically associated differences even when predictive fit is similar. A lightweight CNN--Transformer EEG foundation model pretrained on the Temple University EEG Corpus (TUEG) extracts segment-level representations. Using the Temple University Epilepsy Corpus (TUEP), representations are mapped to a shared latent-state space, and sparse multinomial logistic transition distributions (mLTD) are fit independently to each subject to obtain personalized transition-dependency graphs $W_n$. Analyses include $n{=}198$ subjects (99 epilepsy / 99 non-epilepsy). At $k{=}4$, epilepsy subjects exhibit substantially denser learned dependency structure ($p{=}1.1\times10^{-7}$), with the same pattern at $k{=}6$ (19.90 vs. 13.46; $p{=}5.2\times10^{-5}$). Graph-derived features provide moderate group discrimination under 5-fold subject-wise cross-validation (AUROC 0.68 at $k{=}4$; 0.65 at $k{=}6$). In contrast, held-out log-likelihood is nearly identical between groups at $k{=}4$ ($-0.992$ vs. $-0.991$; $p{=}0.95$), with similarly matched next-state prediction (AUROC 0.855 vs. 0.861; $p{=}0.54$). Thus, similar predictive fit does not imply similar learned dynamics: groups can be comparably predictable while differing substantially in the internal dynamical structure learned by personalized models. This distinction motivates evaluating learned structure alongside predictive performance in personalized clinical models.