Similar Predictive Fit but Different Latent Dynamics: Characterizing Learned Dynamical Structure in Personalized Models of Brain Disorders

Rita Huan-Ting Peng; Nhat Bui

cs.LGOct 7, 2026

Similar Predictive Fit but Different Latent Dynamics: Characterizing Learned Dynamical Structure in Personalized Models of Brain Disorders

Authors: Rita Huan-Ting Peng, Nhat Bui

Organizations: University of Illinois Urbana-Champaign, Urbana, IL, USA · Carle Foundation Hospital, Urbana, IL, USA

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.

Figures & tables

Figure 1 : Pipeline for personalized latent-dynamics modeling from scalp EEG. Non-overlapping 30-s EEG segments are encoded using our frozen EEG foundation model and mapped to shared latent-state spaces using global $k$ -means at $k{=}4$ and $k{=}6$ . Each segment is assigned a discrete state label, producing a subject-specific state sequence. The sequence is modeled independently for each subject using mLTD with group-lasso regularization across temporal lags, yielding a sparse, directed, weighted transition-dependency graph $W_{n}\in\mathbb{R}^{k\times k}$ .

	$k{=}4$			$k{=}6$
Metric	Non	Epi	$p$	Non	Epi	$p$
Predictive fit
Held-out CV LL	$-0.991$	$-0.992$	0.95	$-1.370$	$-1.399$	0.18
Learned dynamical structure
Nonzero dependencies $E_{n}$	5.67	10.01	$1.1{\times}10^{-7}$	13.46	19.90	$5.2{\times}10^{-5}$
Normalized density $D_{n}$	0.354	0.626	–	0.374	0.553	–

Table 1: Similar predictive fit, different learned structure (

n{=}198

; 99 epilepsy / 99 non-epilepsy; Mann–Whitney). Held-out CV LL measures within-subject predictive validity.

E_{n}

denotes the number of retained nonzero dependencies, and

D_{n}=E_{n}/k^{2}

denotes the corresponding normalized dependency density. Because

D_{n}

is a constant rescaling of

E_{n}

within each

k

, separate significance tests are not reported. Mean diagonal strength summarizes self-dependency in

W_{n}

Figure 2 : Main result overview. (A) Epilepsy minus non-epilepsy nonzero-dependency counts across $k$ and lag. (B) Within-subject held-out CV log-likelihood at $k{=}4$ and $k{=}6$ (comparable predictive fit). (C) Group discrimination from $W_{n}$ features (discussed in Section 3.2 ). Dashed line in (C): chance AUROC ( $0.5$ ).

$k$	Feature set	Pooled AUROC	Fold mean $\pm$ SD
4	Scalar summaries	0.697	$0.697\pm 0.058$
4	Graph-derived	0.679	$0.682\pm 0.056$
4	Vectorized $W_{n}$	0.615	$0.617\pm 0.063$
6	Scalar summaries	0.677	$0.685\pm 0.079$
6	Graph-derived	0.645	$0.648\pm 0.111$
6	Vectorized $W_{n}$	0.654	$0.657\pm 0.112$

Table 2: Between-group discrimination from three representations of

W_{n}

: global scalar summaries, graph-derived features, and the vectorized full dependency matrix (stratified 5-fold subject-wise CV;

n{=}198

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 3 : Group-mean transition-dependency matrices. Mean $W_{n}$ at $k{=}4$ (top) and $k{=}6$ (bottom) for non-epilepsy (left) and epilepsy (right). Epilepsy maps are generally denser / stronger, consistent with Table 1 . Latent states are unsupervised and unlabeled.

Figure 4 : Between-group ROC curves from graph-derived $W_{n}$ features (Section 3.2 ). These AUROCs measure epilepsy vs. non-epilepsy discrimination and are distinct from within-subject next-state prediction ( $\approx 0.86$ ).

Figure 5 : Lag checks at $k{=}4$ . (A) Mean $E_{n}$ among subjects with best_lag $=L$ (small subsets; exploratory). (B) Held-out CV log-likelihood vs. lag for all labeled subjects ( $n{=}198$ ; mean $\pm$ SD). Fit remains matched across groups.

Lag $L$	$n$	Epi	Non	Non $E$	Epi $E$	$\Delta E$	$p$
1	16	9	7	5.43	8.11	$+2.68$	0.267
2	13	7	6	7.67	11.29	$+3.62$	0.210
3	18	14	4	9.00	9.21	$+0.21$	0.914
4	11	4	7	8.29	14.25	$+5.96$	0.076
5	7	5	2	11.00	12.00	$+1.00$	1.000
6	14	10	4	13.00	13.40	$+0.40$	0.741

Appendix

Table 3 : Best-lag subset densification at

k{=}4

(exploratory).

E

denotes mean nonzero dependencies;

\Delta E=

epilepsy

-

non-epilepsy (Mann–Whitney

p

Lag $L$	CV LL Non	CV LL Epi	$p$
1	$-1.041$	$-1.034$	0.65
2	$-1.032$	$-1.022$	0.56
3	$-1.033$	$-1.017$	0.37
4	$-1.030$	$-1.018$	0.62
5	$-1.029$	$-1.019$	0.67
6	$-1.030$	$-1.017$	0.55

Appendix

Table 4 : Full-cohort held-out CV log-likelihood vs. lag at

k{=}4

(best regularization at each lag;

n{=}99

per group; Mann–Whitney

p

Explore similar work

NeuroDyn-EEG: An Interpretable Pre-trained Model for EEG Based on Neural Dynamics
Sep 29, 2026Yi Cui, Tong Zhao, Jiaxin Lei +5Electroencephalography Neural Posterior Estimation
Gnosis Neurodynamics Co. Ltd, Room 202, West Area, Building North 1, No. 9 Yard, Shengmingyuan Road, Changping District, Beijing, China · School of Biomedical Engineering, Tsinghua Medicine, Tsinghua University, Beijing, China · University of California, Davis, department of psychology +2
EEG-JEPA: Structured Latent Prediction for EEG Foundation Models
Jul 31, 2026Jinhao Li, Zhiyuan Ma, Xueqiao Han +8Self-Supervised Time Series Representation Learning Joint-Embedding Predictive Architecture
Tsinghua Laboratory of Brain and Intelligence, Tsinghua University · School of Basic Medical Sciences, Tsinghua Medicine, Tsinghua University · School of Biomedical Engineering, Tsinghua Medicine, Tsinghua University +2
What Do EEG Foundation Models Capture from Human Brain Signals?
May 12, 2026Ling Tang, Qian Chen, Jilin Mei +6Frequency-Domain Feature Learning Representation Probing
Shanghai Artificial Intelligence Laboratory · Shanghai Jiao Tong University · Tongji University +2

Sep 29, 2026, cs.NEJ/K move · Enter open · S save

NeuroDyn-EEG: An Interpretable Pre-trained Model for EEG Based on Neural Dynamics

Yi Cui, Tong Zhao, Jiaxin Lei, Chuyi Yang+4

Gnosis Neurodynamics Co. Ltd, Room 202, West Area, Building North 1, No. 9 Yard, Shengmingyuan Road, Changping District, Beijing, China · School of Biomedical Engineering, Tsinghua Medicine, Tsinghua University, Beijing, China · University of California, Davis, department of psychology+2

Electroencephalography Neural Posterior Estimation

Open paper AskPDF arXiv

Clinical scalp electroencephalography (EEG) offers a noninvasive window into neural dynamics of neuropsychiatric disorders. However, discriminative deep models often lack anatomically indexed physiological interpretability. We propose NeuroDyn-EEG, a pretraining framework integrating generative priors from neural dynamics. It couples an extended Jansen-Rit neural mass model, leadfield-based source projection, and simulation-based parameter inversion. Trained on synthetic parameter-EEG pairs within physiological ranges, NeuroDyn-EEG estimates 11 regional parameter families across 90 AAL regions plus one global parameter from standard 19-channel EEG, using only ~2.43M trainable parameters. We evaluate the framework across three levels. First, controlled simulations demonstrate robust parameter recovery under diverse noise conditions, while real resting-state EEG evaluations confirm spectral and phase consistency in an inverse-forward closed loop. Second, on four clinical benchmarks (AD65, PD31, Figshare MDD, and TUAB), NeuroDyn-EEG achieves competitive classification performance, securing the highest BACC, AUROC, and AUCPR on PD31 and MDD, and highest BACC on AD65. Third, post hoc regional analyses reveal disease-specific alterations: local synaptic connectivity C_1 involves the most altered regions in AD65, whereas the firing threshold theta ranks first in MDD, offering testable mechanistic hypotheses. Overall, NeuroDyn-EEG maps scalp EEG to anatomically indexed dynamical parameters, bridging representation learning and mechanistic neurophysiology. Code: https://github.com/Gnosis-Neurodynamics/NeuroDyn-EEG.

Abstract

Explore similar work

NeuroDyn-EEG: An Interpretable Pre-trained Model for EEG Based on Neural Dynamics

EEG-JEPA: Structured Latent Prediction for EEG Foundation Models

What Do EEG Foundation Models Capture from Human Brain Signals?