Investigating the Utility of Explainable Artificial Intelligence for Neuroimaging-Based Dementia Diagnosis and Prognosis.

Comprehensive Summary

This study evaluates whether explainable artificial intelligence (XAI) methods can help validate and improve dementia prediction models that analyze brain MRI scans. Researchers used T1-weighted MRI data from 3,410 participants in the National Alzheimer's Coordinating Centre and 484 from the Alzheimer's Disease Neuroimaging Initiative to train two types of neural network models (ResNets and Vision Transformers) on two tasks: identifying Alzheimer's disease versus healthy controls (diagnosis) and predicting which patients with mild cognitive impairment would progress to dementia (prognosis). The diagnosis models achieved 81 percent balanced accuracy and the prognosis models achieved 67 percent balanced accuracy. The researchers applied eleven different XAI explanation techniques to the trained models to identify which brain regions each model relied upon to make predictions. Gradient-based XAI methods (such as backpropagation and integrated gradients) showed strong agreement in highlighting brain regions relevant to Alzheimer's disease, including the hippocampus, cingulate gyrus, and ventricles. The authors propose a framework for interpreting XAI outputs by treating different explanation methods as independent reviewers that assess model decisions, with agreement across multiple methods strengthening confidence in identified features.

Outcomes and Implications

This work is important because it demonstrates that XAI can serve as a quality control tool to verify that dementia prediction models are focusing on clinically relevant brain regions rather than learning hidden biases. Many current AI diagnostic tools are difficult to trust because their decision-making process remains opaque. The study shows that when saliency maps (visual representations of important regions) from XAI methods were incorporated as features into a separate predictive model for detecting which mild cognitive impairment patients would progress to dementia, prediction accuracy improved by up to 16.7 percent compared to models using only clinical information. However, significant barriers prevent immediate clinical implementation. The models were trained and validated on laboratory brain scans using standard protocols, which may not reflect the variable imaging quality and diverse patient populations seen in routine clinical practice. Additionally, researchers still lack a reliable ground-truth method to confirm whether XAI explanations are truly identifying disease-relevant patterns versus learned artifacts. Before these tools could be used in clinical settings, validation studies in diverse patient populations and prospective clinical trials would be required, likely requiring 3 to 5 years of additional research.