Cross-modal Causal Intervention for Alzheimer's Disease Prediction

Yutao Jin, Haowen Xiao, Jielei Chu, Fengmao Lv, Yuxiao Li, Tianrui LiarXiv 2025 · DOI · OpenReview

Introduction

Alzheimer’s Disease (AD) diagnosis remains a significant challenge in clinical practice. While multi-modal data—such as Magnetic Resonance Imaging (MRI) and clinical records—provide a wealth of information, deep learning models often struggle with "noisy" data. Specifically, these models are susceptible to confounders: variables that correlate with both the input data and the final diagnosis without being the actual cause of the disease. For instance, age-related brain atrophy or unrelated vascular lesions can mimic AD symptoms on an MRI, leading standard models to rely on spurious correlations rather than genuine pathological signals.

The MediAD Framework

To address these challenges, researchers have developed MediAD, a novel visual-language framework that leverages causal inference to improve diagnostic accuracy. MediAD moves beyond simple black-box classification by modeling the diagnostic process as a structural causal graph.

1. Structural Causal Modeling and Front-Door Adjustment

A key innovation of MediAD is the use of Front-Door Adjustment (FDA) to mitigate the impact of unobservable confounders. By introducing a "mediator" variable—an intermediate step influenced by the input which then influences the diagnosis—the model can mathematically "block" the influence of hidden confounding paths. This allows the system to focus on the direct causal effect of the medical data on the diagnosis, even when some confounders are unrecorded.

2. LLM-Enriched Textual Data

MediAD utilizes Large Language Models (LLMs), specifically DeepSeek, to transform unstructured clinical notes and medical histories into structured, high-fidelity summaries. This "knowledge distillation" ensures that the model receives comprehensive textual features (covering basic info, medical history, and daily behavior) to complement raw numerical and visual data.

3. Cross-modal Causal Fusion (CF)

The framework generates the mediator by fusing visual features (extracted via 3D-CNNs or Vision Transformers from MRIs) with the enriched textual features. To ensure the stability of this mediator, the authors applied consistency regularization and employed Monte Carlo sampling to rigorously calculate the mathematical expectation of the causal effect, leading to more reliable interventions than simple attention-based approximations.

Experimental Evaluation

The model was rigorously tested on two major datasets: the NACC (National Alzheimer’s Coordinating Center), containing multi-center data from 39 US research centers, and the ADNI (Alzheimer’s Disease Neuroimaging Initiative), a global dataset for early diagnostic biomarkers.

Key Results

MediAD demonstrated state-of-the-art performance across multiple benchmarks, outperforming models like MedBlip and NeuroSysAD:

  • Multi-class Accuracy: Achieved 90.38% on ADNI and 85.69% on NACC for distinguishing between Healthy Control (CN), Mild Cognitive Impairment (MCI), and AD.
  • Binary Classification: Reached a near-perfect 99.79% AUC for CN vs. AD on the ADNI dataset.
  • Ablation Studies: Removing the LLM-generated summaries led to a significant performance drop (from ~85% to ~81% on NACC), highlighting the importance of textual enrichment.

Conclusion

MediAD represents a significant step forward in applying causal reasoning to medical AI. By prioritizing causality over simple correlation and leveraging the summarization capabilities of LLMs, the framework provides a more robust and "honest" diagnostic tool. Future work will explore more sophisticated mediator architectures to further capture the complexities of neurological disease progression.

Summary updated

Abstract

Mild Cognitive Impairment (MCI) serves as a prodromal stage of Alzheimer's Disease (AD), where early identification and intervention can effectively slow the progression to dementia. However, diagnosing AD remains a significant challenge in neurology due to the confounders caused mainly by the selection bias of multi-modal data and the complex relationships between variables. To address these issues, we propose a novel visual-language causality-inspired framework named Cross-modal Causal Intervention with Mediator for Alzheimer's Disease Diagnosis (MediAD) for diagnostic assistance. Our MediAD employs Large Language Models (LLMs) to summarize clinical data under strict templates, therefore enriching textual inputs. The MediAD model utilizes Magnetic Resonance Imaging (MRI), clinical data, and textual data enriched by LLMs to classify participants into Cognitively Normal (CN), MCI, and AD categories. Because of the presence of confounders, such as cerebral vascular lesions and age-related biomarkers, non-causal models are likely to capture spurious input-output correlations, generating less reliable results. Our framework implicitly mitigates the effect of both observable and unobservable confounders through a unified causal intervention method. Experimental results demonstrate the outstanding performance of our method in distinguishing CN/MCI/AD cases, outperforming other methods in most evaluation metrics. The study showcases the potential of integrating causal reasoning with multi-modal learning for neurological disease diagnosis.
BibTeX
@misc{jin2025cross,
  author = {Yutao Jin and Haowen Xiao and Jielei Chu and Fengmao Lv and Yuxiao Li and Tianrui Li},
  title = {{Cross-modal Causal Intervention for Alzheimer's Disease Prediction}},
  year = {2025},
  eprint = {2507.13956},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.13956},
}
Figure