LearnAD: Learning Interpretable Rules for Brain Networks in Alzheimer's Disease Classification

TL;DR

The paper tackles the interpretability gap in Alzheimer's disease prediction from structural MRI by introducing LearnAD, a neuro-symbolic framework that combines statistical feature selection (DT, RF, or GCN) with symbolic rule learning via FastLAS to produce compact, human-readable rules. By generating context-rich examples from selected brain connections and leveraging domain knowledge, LearnAD yields globally interpretable rules that achieve competitive accuracy compared to non-interpretable baselines, while offering far greater transparency. Ablation studies show the neuro-symbolic approach improves DT performance and approaches RF/GNN accuracy on full feature sets, with a substantial interpretability advantage (rules with far fewer atoms). The approach provides mechanistic biomarkers (e.g., left temporal pole–left hippocampus, right precuneus–right superior parietal) and offers a framework for extending to time-series data and multi-modal imaging, enhancing clinical trust and understanding of network-level disease progression.

Abstract

We introduce LearnAD, a neuro-symbolic method for predicting Alzheimer's disease from brain magnetic resonance imaging data, learning fully interpretable rules. LearnAD applies statistical models, Decision Trees, Random Forests, or GNNs to identify relevant brain connections, and then employs FastLAS to learn global rules. Our best instance outperforms Decision Trees, matches Support Vector Machine accuracy, and performs only slightly below Random Forests and GNNs trained on all features, all while remaining fully interpretable. Ablation studies show that our neuro-symbolic approach improves interpretability with comparable performance to pure statistical models. LearnAD demonstrates how symbolic learning can deepen our understanding of GNN behaviour in clinical neuroscience.

Figures (3)

• Figure 1: Architecture of our neuro-symbolic approach. The model consists of 4 components: Statistical ML Model, Feature Selector, Network-to-Knowledge Generator, and Symbolic Model. The dataset $(X,y)$ is provided to the model, along with the background knowledge, $B$. $\theta$ defines the learned parameters of the statistical models, $X^{(F)}$ are the learned relevant features. $S_M$ is the hypothesis space and $E$ are the examples. The learned hypothesis, $H$, is the output.
• Figure 2: Representative rules learned by LearnAD(DT).
• Figure 3: Inductive bias used in a FastLAS task (examples redacted for privacy).