EVA-Net: Interpretable Anomaly Detection for Brain Health via Learning Continuous Aging Prototypes from One-Class EEG Cohorts

TL;DR

This work tackles EEG-based brain age estimation under imperfect data by reframing it as an anomaly-detection problem. It introduces EVA-Net, a framework that combines an efficient long-sequence encoder with ProbSparse attention, a Variational Information Bottleneck, and an age-conditioned continuous prototype network to learn a robust, interpretable healthy aging manifold. EVA-Net achieves state-of-the-art brain age prediction on healthy subjects and demonstrates effective anomaly detection on unseen MCI/AD patients via the Brain-Age Gap ($$BAG$$) and Prototype Alignment Error ($$PAE$$). The approach offers a principled, interpretable, and scalable pathway for healthcare-grade brain health monitoring using routinely collected EEG data.

Abstract

The brain age is a key indicator of brain health. While electroencephalography (EEG) is a practical tool for this task, existing models struggle with the common challenge of imperfect medical data, such as learning a ``normal'' baseline from weakly supervised, healthy-only cohorts. This is a critical anomaly detection task for identifying disease, but standard models are often black boxes lacking an interpretable structure. We propose EVA-Net, a novel framework that recasts brain age as an interpretable anomaly detection problem. EVA-Net uses an efficient, sparsified-attention Transformer to model long EEG sequences. To handle noise and variability in imperfect data, it employs a Variational Information Bottleneck to learn a robust, compressed representation. For interpretability, this representation is aligned to a continuous prototype network that explicitly learns the normative healthy aging manifold. Trained on 1297 healthy subjects, EVA-Net achieves state-of-the-art accuracy. We validated its anomaly detection capabilities on an unseen cohort of 27 MCI and AD patients. This pathological group showed significantly higher brain-age gaps and a novel Prototype Alignment Error, confirming their deviation from the healthy manifold. EVA-Net provides an interpretable framework for healthcare intelligence using imperfect medical data.

Figures (5)

• Figure 1: The overall architecture of our proposed EVA-Net. EEG signals are converted via Temporal Embedding into a sequence. An efficient backbone (using ProbSparse attention by identifying Top KL time-points) extracts a hidden state $H_i$. This state is passed through an Information Bottleneck (VIB) to produce a robust latent code $Z_i$. $Z_i$ is then used for both Brain Age prediction (regression) and alignment with an ideal age-matched Prototype $P_{y_i}$.
• Figure 2: ProbSparse attention mechanism achieving $O(L \log L)$ complexity. Queries are ranked by KL divergence scores, with top-scoring queries (pink) forming sparse matrix $\bar{Q}$. Computing $\bar{Q}K^T$ bypasses full $L \times L$ attention, reducing computational cost while preserving critical temporal dependencies.
• Figure 3: Age distribution of the normative healthy cohort. A bar chart illustrating the sample count by decade. The distribution is concentrated in adulthood, and fewer subjects in the younger and older brackets, with a cohort mean age of 44.3 years and a median age of 44 years.
• Figure 4: Sensitivity analysis of EVA-Net. (a, b) MAE and RMSE comparison across encoder backbones. (c) VIB loss weight $\beta$ sensitivity within {0.01, 0.02, 0.03, 0.04, 0.05, 0.06}. (d) Prototype alignment loss weight $\gamma$ sensitivity within {0.1, 0.3, 0.5, 0.7, 0.9}, optimal at $\gamma=0.7$.
• Figure 5: Violin plots illustrating the distribution of anomaly scores across cohorts. (a) Brain-Age Gap (BAG): While the healthy control group (N=130) is centered around zero, the MCI (N=12) and AD (N=15) groups show progressively larger positive gaps, indicating accelerated brain aging. (b) Prototype Alignment Error (PAE): A similar trend is observed for PAE, where pathological groups exhibit significantly larger distances from the normative manifold. The violin shapes visualize the full density of the data, highlighting the clear separation between healthy and pathological distributions.