TLDR: A new AI framework, ADPC, uses MRI/fMRI images and LLM-summarized clinical data to predict Alzheimer’s disease (AD), Mild Cognitive Impairment (MCI), and Cognitively Normal (CN) states. It employs causal intervention to eliminate confounding factors, leading to more reliable and accurate diagnoses, achieving state-of-the-art performance on ADNI and NACC datasets.
Alzheimer’s Disease (AD) is a progressive condition that affects memory and thinking, often preceded by Mild Cognitive Impairment (MCI). Early detection is crucial because interventions during the MCI stage can significantly slow down the disease’s progression. However, diagnosing AD is challenging due to various confounding factors, such as biases in multi-modal data and complex relationships between different variables.
Researchers have developed a new framework called Alzheimer’s Disease Prediction with Cross-modal Causal Intervention (ADPC) to assist in diagnosis. This innovative approach tackles the diagnostic challenges by integrating visual and language data with a focus on causal relationships.
The ADPC framework utilizes advanced technologies, including Magnetic Resonance Imaging (MRI) and functional MRI (fMRI) scans, alongside textual data. A key aspect of this framework is its use of large language models (LLMs) to process and summarize clinical data. These LLMs take unstructured clinical information and convert it into a standardized, structured text format, even when the original data is incomplete or unevenly distributed. This ensures that the textual input is consistent and high-quality.
One of the core strengths of ADPC lies in its ability to implicitly eliminate confounders. Confounders are factors that can distort the true relationship between variables, leading to inaccurate predictions. For instance, age-related brain changes or artifacts in neuroimaging could be mistakenly identified as AD-specific features. Traditional models, which are purely data-driven, often capture these spurious correlations. ADPC addresses this by employing a concept called causal intervention, specifically using a “front-door adjustment” mechanism. This involves introducing a “mediator” variable that helps to isolate the true causal effects between the input data (like brain scans and clinical summaries) and the diagnostic outcome, thereby improving reliability.
The framework processes visual features from MRI/fMRI scans using a specialized feature extractor and textual features from the LLM-generated summaries. These features are then combined in a “Cross-modal Causal Fusion” module to create the mediator. This mediator, along with the combined multi-modal features, undergoes the front-door adjustment to remove the influence of confounding factors. Finally, these refined features are fed into a classifier to categorize participants into Cognitively Normal (CN), Mild Cognitive Impairment (MCI), or Alzheimer’s Disease (AD).
Experimental results have shown that ADPC achieves state-of-the-art performance in distinguishing between CN, MCI, and AD cases. The method was tested on two major Alzheimer’s research databases: the National Alzheimer’s Coordinating Center (NACC) and Alzheimer’s Disease Neuroimaging Initiative (ADNI) datasets. Even with a smaller training group on the ADNI dataset compared to some previous methods, ADPC demonstrated superior accuracy, F1-score, precision, recall, and AUC metrics for three-class classification. It also showed excellent performance in binary classification (AD vs. CN).
An ablation study confirmed the importance of the causal fusion module and front-door adjustment, as removing them led to a significant drop in performance. Visualization techniques, such as t-SNE, illustrated clear separation between the different diagnostic categories, indicating the model’s strong discriminative power. Furthermore, the model’s attention analysis revealed that it prioritizes neurological assessment results for AD patients, medical history for CN and MCI individuals, and considers age, smoking, and alcohol consumption as important indicators, aligning with existing medical understanding. Interestingly, the model also showed heightened attention to the term “female” when diagnosing AD in women, consistent with research suggesting higher susceptibility in women.
Also Read:
- AI Models Offer New Insights into Bridge Health Through Non-Destructive Evaluation
- Generative AI Enhances Realism in Human Motion Simulation
In conclusion, the ADPC framework represents a significant step forward in Alzheimer’s disease diagnosis. By explicitly modeling causal relationships and mitigating the impact of confounders through causal intervention, it offers a more reliable and accurate diagnostic tool, paving the way for earlier and more effective interventions. For more details, you can refer to the full research paper.
