🤖 AI

LERD: Latent Event-Relational Dynamics for Neurodegenerative Classification

The paper proposes LERD, an end-to-end Bayesian latent event-relational dynamical system that infers unannotated neural events and their cross-channel interactions from multichannel EEG to improve the accuracy and physiological interpretability of Alzheimer's disease classification.

Original authors: Yicheng Feng, Hairong Chen, Ziyu Jia, Samir Bhatt, Hengguan Huang

Published 2026-06-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yicheng Feng, Hairong Chen, Ziyu Jia, Samir Bhatt, Hengguan Huang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your brain is like a bustling city with thousands of tiny radio stations (neurons) broadcasting signals. In a healthy city, these stations have a rhythm: they broadcast at specific times, and the stations often coordinate with each other, like a well-orchestrated symphony.

In Alzheimer's disease, this city gets chaotic. The radio stations start broadcasting at the wrong times, the rhythm slows down, and the coordination between stations breaks. Doctors use EEG machines to listen to these radio signals from the scalp, but the signals are messy, noisy, and mixed together.

The Problem with Current Tools
Most current computer programs used to diagnose Alzheimer's are like "black box" judges. They listen to the noisy radio signals and say, "This person has Alzheimer's" or "This person is healthy." They are good at getting the right answer, but they don't explain why. They can't tell you when the radio stations went off or how the stations stopped talking to each other. It's like a judge giving a verdict without showing the evidence.

The Solution: LERD
The authors of this paper created a new tool called LERD. Think of LERD not as a judge, but as a detective with a time machine.

Instead of just guessing the final diagnosis, LERD tries to reconstruct the hidden story behind the noise. It asks two big questions:

When did the hidden "events" (the brain's internal signals) actually happen?
How did the different parts of the brain influence each other in the moments leading up to those events?

How LERD Works (The Analogy)
To understand how LERD works, imagine you are trying to figure out the schedule of a busy train station, but you only have a recording of the crowd's noise and no train schedules.

The "Event Detective" (EPDE): LERD looks at the messy noise and tries to pinpoint exactly when a "train" (a neural event) left the station. It doesn't just guess; it uses a special mathematical rulebook (called a "dLIF prior") that acts like a physics teacher. This teacher says, "Trains can't leave every millisecond; they need to recharge, and they have a maximum speed." This keeps the detective's guesses realistic and grounded in how brains actually work.
The "Time Traveler" (MELP): Once the detective spots a train leaving, LERD predicts when the next train will come. It knows that sometimes trains come quickly, and sometimes there's a long wait. It uses a flexible "log-normal" system to guess these time gaps, allowing for a wide variety of realistic schedules rather than a rigid, robotic one.
The "Connection Map" (ERG): Finally, LERD draws a map showing how the stations talk to each other. If Station A sends a signal, does Station B respond 0.1 seconds later? Or does it ignore it? LERD builds a dynamic map of these connections, showing who is influencing whom and when.

What They Found
The researchers tested this detective tool in two ways:

The Fake City (Synthetic Data): They created a fake brain with known rules and hidden events. They asked other computer programs to find the hidden events. The other programs were great at guessing the final answer but failed completely at finding the hidden events (their "IoU" score was zero). LERD, however, successfully found the hidden events and the timing, proving it could actually see the structure behind the noise.
The Real City (Real Alzheimer's Data): They tested LERD on real EEG recordings from patients with Alzheimer's, other types of dementia, and healthy people.
- Accuracy: LERD was better at correctly identifying who had Alzheimer's compared to other top methods.
- Stability: Unlike other tools that sometimes guessed right and sometimes wrong depending on the day, LERD was consistent.
- Insight: Most importantly, LERD produced summaries that matched what neuroscientists already know about the disease (like slowing brain rhythms), but it did so by discovering these patterns directly from the data without being told what to look for.

The Bottom Line
LERD is a new way to listen to the brain. It doesn't just tell you if a patient is sick; it reconstructs the hidden timeline of their brain's activity, showing when things went wrong and how the brain's communication network fell apart. It turns a black-box prediction into a transparent, physics-based story of the brain's dynamics.

Technical Summary: LERD – Latent Event-Relational Dynamics for Neurodegenerative Classification

1. Problem Formulation

The paper addresses the challenge of diagnosing Alzheimer's disease (AD) and characterizing neurodegenerative dynamics using multichannel electroencephalography (EEG). While deep learning has advanced EEG-based classification, existing approaches typically function as "black-box" label predictors. They map spatiotemporal signals to diagnostic categories (e.g., AD vs. control) without explicitly modeling the latent neural events (population-level transitions) or the temporal coordination (cross-channel interactions) that generate the signals.

The authors identify two primary gaps:

Lack of Latent Recovery: Current models do not recover the timing of latent events or the directed interaction graphs between channels, making it difficult to relate predictions to established neurophysiological phenomena (e.g., event-driven activity patterns or timing-derived interaction summaries).
Annotation Scarcity: Inferring these dynamics is difficult because scalp EEG is a noisy, linear mixture of mesoscopic sources, and the specific event times and interaction edges are unobserved (latent) and unannotated.

The objective is to design a system that simultaneously infers:

Channel-wise latent event dynamics (when events occur).
A conditional event-relational graph (how events in one channel probabilistically influence others).
Diagnostic labels, without requiring spike or edge annotations.

2. Methodology: The LERD Framework

LERD (Latent Event-Relational Dynamics) is an end-to-end Bayesian neural dynamical system trained via variational inference. It consists of three core interacting components:

A. Event Posterior Differential Equation (EPDE)

The EPDE models the posterior distribution of latent event times for each channel given the observed EEG signal.

Instead of predicting discrete time steps, it defines an auxiliary function $\Phi(t)$ whose derivative accumulates the contribution of the event density.
The expected next event time is computed as the solution to an Initial Value Problem (IVP), ensuring events remain strictly ordered in time.
This provides a continuous-time representation of latent dynamics and expected next-event times.

B. Mean–Evolving Lognormal Process (MELP)

To capture the stochastic nature of inter-event intervals, LERD employs a MELP.

The EPDE outputs serve as mean parameters for a mixture of lognormal distributions.
Using reparameterized sampling (e.g., Gumbel-Softmax during training), the model draws stochastic inter-event intervals.
This allows the model to handle flexible, potentially multimodal timing statistics while ensuring positive intervals.

C. Event-Relational Graph (ERG)

The ERG captures directed interactions between channels based on the inferred event lags.

Cross-channel event lags are mapped through a smooth nonlinearity (inspired by Spike-Timing-Dependent Plasticity, STDP) into edge weights.
The graph is not a static connectivity matrix but a dynamic process summarizing how events in one channel precede, follow, or align with events in others.
A weak regularizer aligns the ERG with observable statistics (e.g., Pearson correlations) but relies primarily on the inferred event lags for fine-grained structure.

D. Electrophysiology-Informed Prior (dLIF)

A key innovation is the incorporation of a differentiable Leaky-Integrate-and-Fire (dLIF) prior.

This imposes biophysically meaningful constraints on the latent process, including leak, refractory behavior, and plausible firing rates.
Since the hard spike function in standard LIF models is non-differentiable, LERD regularizes the rate proxies derived from the EPDE state to follow dLIF laws.
This acts as a structured prior on population-level latent rate dynamics rather than a claim that scalp EEG reflects single-neuron spikes.

E. Training Objective

The model is trained by minimizing a negative Evidence Lower Bound (ELBO) objective:
$J = -\mathbb{E}[\log p_\theta(Y|X, T, G)] + \text{KL}(T) + \text{KL}(\tau) + \beta R_{\text{ERG}} + \lambda_{\text{LIF}} R_{\text{LIF}}$

$\text{KL}(T)$ : A tractable IVP-based surrogate for the KL divergence between the inferred path law and the dLIF prior.
$\text{KL}(\tau)$ : Penalizes deviation from a lognormal mixture prior for inter-event intervals.
$R_{\text{ERG}}$ & $R_{\text{LIF}}$ : Weak regularizers for graph alignment and dLIF consistency, respectively.

3. Key Contributions

Annotation-Free Discovery: LERD is a unified Bayesian system that infers latent event processes and event-relational graph dynamics directly from multichannel EEG without requiring spike or edge annotations.
Biophysical Constraints: It integrates a differentiable dLIF prior, providing constraints on rates, refractoriness, and frequency content that align with electrophysiological principles.
Theoretical Guarantees: The authors derive a computationally tractable IVP-based representation for the event-prior KL and establish stability bounds for the inferred ERG with respect to lag noise.
Empirical Performance: The model demonstrates the ability to recover latent structures that align with physiological differences while achieving competitive diagnostic performance.

4. Experimental Results

Synthetic Benchmarks (Toy Dataset)

Setup: Synthetic data with known latent event rates and boundaries across three frequency bands.
Comparison: Compared against Neural ODEs (NODE), ODE-RNN, and STRODE.
Findings: While baselines achieved high cosine similarity (sequence prediction), they failed to recover latent event structures (Intersection-over-Union (IoU) scores were 0.000).
LERD Performance: Achieved high cosine similarity while obtaining non-zero IoU scores (e.g., 0.473 in the 5–10 Hz band), demonstrating its ability to recover the underlying generative mechanisms and event rates.

Real-World AD EEG Datasets

Datasets: Two cohorts (A and B) containing AD, Frontotemporal Dementia (FTD), Mild Cognitive Impairment (MCI), and healthy controls.
Performance: LERD outperformed strong baselines (CNNs, RNNs, Transformers) in diagnostic accuracy.
- In Cohort A, it showed clear distinction between AD, FTD, and controls with lower performance variability.
- In the more unbalanced Cohort B, it maintained robustness and high accuracy where other models fluctuated.
Interpretability: The model yielded "physiology-aligned" summaries (rates, timing, and graph structures) that helped characterize group-level dynamical differences, such as oscillatory slowing and disrupted coordination.

5. Significance and Claims

The paper claims that LERD bridges the gap between high-performance black-box classification and mechanistic neuroscientific understanding. By explicitly modeling latent event timing and event-relational dynamics, the framework:

Provides a principled way to infer directed, event-level interactions from noisy, unannotated EEG data.
Offers interpretable summaries (event rates, timing summaries, and interaction graphs) that can be compared against established EEG phenomena.
Achieves superior diagnostic performance while maintaining alignment with electrophysiological constraints, suggesting that incorporating domain knowledge (via the dLIF prior) improves both the robustness and the interpretability of deep learning models in neurodegenerative classification.

The authors emphasize that the dLIF prior is phenomenological, serving as a structured regularizer for population-level dynamics rather than a direct mapping to single-neuron physiology. The work positions itself as a step toward "mechanistic Bayesian reasoning" in EEG analysis, moving beyond static feature extraction to dynamic, event-driven modeling.