Similar Predictive Fit but Different Latent Dynamics: Characterizing Learned Dynamical Structure in Personalized Models of Brain Disorders
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 Wn. Analyses include n=198 subjects (99 epilepsy / 99 non-epilepsy). At k=4, epilepsy subjects exhibit substantially denser learned dependency structure (p=1.1×10−7), with the same pattern at k=6 (19.90 vs. 13.46; p=5.2×10−5). 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.
Related
- Deciphering plant stress resilience in different millet crops: from multi-omics discoveries to predictive breeding for climate-resilient crop enhancement
- Structure-Aware Mapping with Dynamical Sources for Informer Weight Initialization in Time-Series Forecasting
- Effects of straw return combined with different amendments on soil enzyme activities and bacterial community structure in acidic purple soil
- Reward changes the use and structure of predictive spatial maps
- Hypothesis-Driven Identification of Neural Algorithms With Dynamical Structure-Preserving Manifolds.
BibTeX
@misc{peng2026similar,
title = {{Similar Predictive Fit but Different Latent Dynamics: Characterizing Learned Dynamical Structure in Personalized Models of Brain Disorders}},
author = {Rita Huan-Ting Peng and Nhat Bui},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.10850},
eprint = {2610.10850},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.10850},
}
RIS
TY - UNPB
TI - Similar Predictive Fit but Different Latent Dynamics: Characterizing Learned Dynamical Structure in Personalized Models of Brain Disorders
AU - Rita Huan-Ting Peng
AU - Nhat Bui
PY - 2026
DA - 2026/10/07
PB - arXiv
DO - 10.48550/arxiv.2610.10850
UR - https://arxiv.org/abs/2610.10850
AB - 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=1.1×10^−7), with the same pattern at k=6 (19.90 vs. 13.46; p=5.2×10^−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.
ER -