Brain Latent Progression (BrLP) Models

Updated 28 January 2026

Brain Latent Progression (BrLP) is a set of quantitative frameworks that model the temporal evolution of brain anatomy, pathology, or function using low-dimensional latent variables.
BrLP integrates longitudinal neuroimaging and biomarker data with probabilistic models and deep learning to generate precise, individualized forecasts for neurodegenerative diseases and brain aging.
Recent advances, such as latent diffusion processes and Bayesian trajectory models, enhance prediction accuracy and interpretability with reductions in volumetric error and increases in SSIM.

Brain Latent Progression (BrLP) refers to a family of quantitative frameworks for modeling, predicting, and interpreting the individualized spatiotemporal evolution of brain anatomy, pathology, or function in terms of low-dimensional latent variables or trajectories. BrLP methodologies integrate large-scale longitudinal neuroimaging or biomarker data into a coherent latent space, within which disease and developmental dynamics can be modeled, forecast, or even generatively simulated at the subject level. Modern BrLP variants employ probabilistic models, latent diffusion processes, and deep learning architectures to achieve precise, temporally consistent, and patient-specific predictions of future brain states, with principal application domains in neurodegenerative diseases and brain aging (Puglisi et al., 12 Feb 2025, Puglisi et al., 2024, Kapoor et al., 26 Aug 2025, Litrico et al., 2024, Litrico et al., 3 Sep 2025, Zhu et al., 2018, Marinescu et al., 2019).

1. Conceptual Foundations and Theoretical Formulations

Early formulations of Brain Latent Progression were grounded in the notion that brain- or disease-related processes unfold along unobservable (latent) temporal or structural axes. In generative disease-progression models, the latent trajectory is variously parameterized as:

A subject-invariant "disease clock" or progression time with subject-specific onset and progression rates, as in DIVE (Marinescu et al., 2019) and Lespinasse et al. (Lespinasse et al., 2023).
A multi-modal low-dimensional latent state evolving nonlinearly or as a Markov process, giving rise to observed multi-domain biomarker or imaging values (Zhu et al., 2018, Cai et al., 2024).
A stochastic process in a D-dimensional latent space describing spatiotemporal deformation or intensity change, as in latent-diffusion-based BrLP models (Puglisi et al., 12 Feb 2025, Puglisi et al., 2024).

The typical mathematical form posits, for observed data $y_{ij}^k$ (measurements of subject $i$ , modality $k$ at time $j$ ), a generative path:

$\text{Latent progression parameters (e.g., time-shift %%%%4%%%%, speed %%%%5%%%%)} \implies \text{Latent state} \implies \text{Observation model}$

For imaging, these latent states may correspond to encoded representations of MRIs, or even intensity-difference maps between timepoints (Litrico et al., 2024, Puglisi et al., 12 Feb 2025). For biomarkers, they can parameterize sigmoid or monotone temporal trajectories (Zhu et al., 2018, Marinescu et al., 2019).

2. Diffusion-based and Latent-space BrLP Models

Recent advances in BrLP leverage latent diffusion models (LDMs) to address scalability, expressivity, and stochasticity in full 3D neuroimaging datasets. The workflow consists of several core modules (Puglisi et al., 12 Feb 2025, Puglisi et al., 2024):

Autoencoder (AE): Encodes $x \in \mathbb{R}^{H \times W \times D}$ (brain MRI) into a compact latent tensor $z_0$ (e.g., $3 \times 16 \times 20 \times 16$ ), and provides a decoder for image reconstruction.
Latent Diffusion Model: Trains a UNet denoiser $\epsilon_\theta$ to predict noise in the noising process $z_t = \sqrt{\bar\alpha_t} z_0 + \sqrt{1-\bar\alpha_t} \epsilon$ , with $i$ 0.
ControlNet Conditioning: Baseline latent codes (baseline anatomy) are introduced via parallel cross-attention modules, ensuring the trajectory preserves individual anatomical features during generation.
Auxiliary Disease Model ( $i$ 1): Predicts future atrophy or biomarker values (e.g., region volumes) given subject metadata and prior values, through linear regression or logistic disease-course mapping.

At inference, stochastic draws from the diffusion prior are averaged using the Latent Average Stabilization (LAS) algorithm, which reduces spatiotemporal noise and quantifies uncertainty.

BrLP models explicitly integrate subject metadata, longitudinal covariates, and prior estimates of progression-related volumes, enhancing individual-level predictions. Conditioning occurs both in the latent-diffusion backbone and via concatenation or modulation in the ControlNet. Auxiliary models—such as small MLPs for single-scan settings, or Disease Course Mapping for longitudinal series—predict future region measures used in conditional generation (Puglisi et al., 12 Feb 2025, Puglisi et al., 2024).

3. Linear and Bayesian Latent-space Trajectories

MRExtrap (Kapoor et al., 26 Aug 2025) and other linear BrLP models posit that, after training a dedicated autoencoder, subject trajectories in latent space are approximately linear with respect to age. For a subject with latent code $i$ 2 at age $i$ 3, the predicted future latent at age $i$ 4 is:

$i$ 5

The progression rate $i$ 6 can be set as a global empirical average, a subject-specific estimate, or a Bayesian posterior updated with previous scans. Empirical evidence demonstrates that these trajectories yield affine changes in regional volumes and standard atrophy measures, distinguishing disease states and accurately stratifying progression (Kapoor et al., 26 Aug 2025).

4. Temporally-/Anatomically-Aware and Identity-Preserving Extensions

Subsequent models emphasize:

Temporal awareness: BrLP models such as TADM and TADM-3D (Litrico et al., 2024, Litrico et al., 3 Sep 2025) integrate explicit age-gap embeddings (sinusoidal positional encoding), and employ brain-age estimators (BAE) as auxiliary losses enforcing the generated MRIs correspond to target ages. Bidirectional temporal regularization further encourages true causal—rather than interpolative—progression.
Anatomical guidance: AG-LDM (Wan et al., 21 Jan 2026) unifies segmentation-based supervision with diffusion training, enhancing morphometric fidelity and counterfactual plausibility, and surpasses BrLP in both image quality (MSE = 0.003 vs. 0.005–0.006) and regional volume MAE.
Identity preservation: IP-LDM (Huang et al., 11 Mar 2025) introduces triplet contrastive losses and identity control nets to maintain subject identity during age transformations and progression modeling, achieving higher structural similarity (SSIM up to 0.949) than previous approaches.

5. Evaluation Protocols and Empirical Results

BrLP models are typically evaluated using:

Image similarity metrics: MSE, SSIM, PSNR, and perceptual scores (e.g., FID, LPIPS).
Volumetric accuracy: MAE of predicted vs. real region volumes (hippocampus, amygdala, ventricles, thalamus, cortex, WM, CSF), often normalized as % total brain volume (Puglisi et al., 12 Feb 2025, Puglisi et al., 2024, Wan et al., 21 Jan 2026, Kapoor et al., 26 Aug 2025).
Downstream classification: Latent-space features are used for disease staging, progression prediction, and discrimination between diagnostic groups (Puglisi et al., 12 Feb 2025, Xu et al., 2020).

Key results include:

18–25% reduction in volumetric MAE and up to 43% increase in SSIM for BrLP compared to GAN or VAE-based methods (Puglisi et al., 12 Feb 2025, Puglisi et al., 2024).
TADM yields a 24% lower region volume error and a 4% gain in SSIM (Litrico et al., 2024).
Global anatomical errors (WASABI) in BrLP are higher than in anatomically-guided models; sensitivity to covariates (e.g., starting/follow-up age) is limited (4–16% MSE change when removed), with AG-LDM showing 25–31× greater dependence (Wan et al., 21 Jan 2026).

6. Theoretical Guarantees, Uncertainty, and Interpretability

Several BrLP frameworks provide theoretical and practical guarantees:

Consistency/identifiability: Hidden Markov and discriminative EM BrLP approaches demonstrate consistency, oracle variable selection, and identifiability under reasonable regularity conditions (Cai et al., 2024).
Uncertainty quantification: LAS in BrLP and posterior sampling in MRExtrap provide variance-based uncertainty measures, correlating with prediction distance and model error (Puglisi et al., 12 Feb 2025, Kapoor et al., 26 Aug 2025).
Interpretability: Linear and logistic-curve-based models permit direct reading of disease timing, inflection points, and the influence of risk factors (e.g., APOE, education, sex) on progression rates (Zhu et al., 2018, Lespinasse et al., 2023).

BrLP frameworks often support retrospective alignment of marker trajectories on latent disease time axes, disambiguating temporal order among multimodal changes (e.g., tau accumulation, imaging atrophy, cognitive impairment) (Lespinasse et al., 2023).

7. Limitations and Open Challenges

Recognized limitations include:

Bias towards healthy aging: All methods, including advanced BrLP variants, show reduced accuracy in AD subjects relative to controls, suggesting incomplete disentanglement of disease-specific vs. normative trajectories (Puglisi et al., 12 Feb 2025, Wan et al., 21 Jan 2026).
Restricted anatomical supervision: Most BrLP models lack explicit global segmentation loss, limiting morphometric plausibility outside optimized ROIs (Wan et al., 21 Jan 2026).
Incomplete integration of multimodal or non-imaging covariates (e.g., genetics, fluid biomarkers) in most pipelines.
Simple parametric trajectory (linear or sigmoid) assumptions in some BrLP approaches may not capture strongly nonlinear, region- or subject-specific deviations, especially in advanced or heterogeneous disease (Kapoor et al., 26 Aug 2025, Lespinasse et al., 2023).

Ongoing research aims at hierarchical or non-parametric latent trajectory modeling, full end-to-end clinical-to-imaging conditioning, robust handling of missing data, and domain generalization across diseases and institutions. Applications include clinical-trial patient selection, forecasting, digital twin generation, and plausibility screening for image syntheses (Puglisi et al., 12 Feb 2025, Puglisi et al., 2024, Wan et al., 21 Jan 2026).

References:

(Puglisi et al., 12 Feb 2025, Puglisi et al., 2024, Kapoor et al., 26 Aug 2025, Wan et al., 21 Jan 2026, Litrico et al., 2024, Litrico et al., 3 Sep 2025, Huang et al., 11 Mar 2025, Cai et al., 2024, Zhu et al., 2018, Marinescu et al., 2019, Lespinasse et al., 2023, Zhao et al., 2021, Xu et al., 2020)