Latent Dynamics · 2026-10-09
Similar Predictive Fit but Different Latent Dynamics: Characterizing Learned Dynamical Structure in Personalized Models of Brain Disorders
Rita Huan-Ting Peng, Nhat Bui
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
The authors study whether personalized latent dynamics differ between epilepsy and non-epilepsy when predictive fit is similar. Using frozen EEG foundation representations and subject-specific mLTD transition-dependency graphs, they find epilepsy models retain denser dependency structure (e.g., nonzero dependencies and densities) at k=4 and k=6, while held-out log-likelihood and within-subject next-state AUROC are nearly matched.
Why it matters
The authors show a fit–structure dissociation: comparable predictive validity does not imply comparable learned latent dynamical structure in personalized clinical models. They further report that graph-derived features provide moderate epilepsy vs. non-epilepsy discrimination even when predictive fit is similar, motivating evaluation of learned structure alongside predictive performance before intervention-aware patient world modeling.
Method
- Pretrain a lightweight CNN–Transformer EEG encoder on TUEG with self-supervised reconstruction; freeze it for TUEP analyses.
- Map TUEP segment representations (30-s segments) into a shared latent-state space via global k-means (k=4 primary; k=6 for robustness).
- For each subject, fit sparse multinomial logistic transition distributions (mLTD) to estimate a personalized sparse directed latent transition-dependency graph W_n; compare predictive validity vs. learned dynamical structure across groups.
Limitation
Limitations include: learned k-means states are unsupervised with no established clinical/physiological semantics; the analysis is observational case–control so differences in W_n represent predictive temporal dependencies rather than causal mechanisms/treatment effects; non-null sensitivity analyses suggest density differences are not solely due to all-zero graphs; only one frozen EEG foundation model is used; cross-encoder concordance for W_n is not established; prospective test–retest and longitudinal analyses are needed to establish stability and clinical generalizability.
Abstract (from arXiv)
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.
Related papers
- TEX-MEX: Similarity-Guided Temporal Contrastive Learning for Time Series
- Controlled Evaluation of EEG Foundation Models: Random Initialisation, Band-Power and Validation-Selected Probes
- Forecasting Patient Trajectories from the Electronic Health Records with Action-Conditioned World Models
- Learning Intrinsic Dynamics in Latent World Models
- SPERA: Spherical Prior EEG Foundation Model with Geometry- and Frequency-Aware Latent Prediction
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-07 · Indexed by Pub-AI: 2026-10-09 UTC
Similar Predictive Fit but Different Latent Dynamics: Characterizing Learned Dynamical Structure in Personalized Models of Brain Disorders
Preprint posted: 2026-10-07 · Original record ↗
- arxiv
- 2610.10850
Explicit source versions (1)
- v1 · 2026-10-07
- Preprint available
Source date: 2026-10-07 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
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.