Explainable artificial intelligence for the diagnosis of neurodegenerative diseases: a systematic review of methods, applications, challenges, and future directions
TL;DR
Analysis of the state-of-the-art on the use of Explainable Artificial Intelligence for the diagnosis of neurodegenerative diseases, focusing on AI models, explainability methods, datasets, validation methods, clinical applications, and implementation challenges showed that convolutional neural networks and hybrid deep learning models are more widely studied than other methods.
Abstract
The progressive nature and overlapping clinical features of neurodegenerative diseases like Alzheimer's disease, Parkinson's disease, Frontotemporal Dementia, Huntington's disease, Dementia with Lewy Bodies and Amyotrophic lateral sclerosis make them tough to diagnose. While AI has shown great accuracy in diagnosis, it has the drawback of being hard for users to interpret, making it difficult to be broadly adopted in clinical practice. The current review aims at analyzing the state-of-the-art on the use of Explainable Artificial Intelligence (XAI) for the diagnosis of neurodegenerative diseases, focusing on AI models, explainability methods, datasets, validation methods, clinical applications, and implementation challenges. The review was conducted according to PRISMA 2020 guidelines and included the PubMed, IEEE Xplore and ScienceDirect databases of research articles published from 2018 to 2026. A total of 25 studies were included for qualitative synthesis from 454 retrieved studies. The results showed that convolutional neural networks and hybrid deep learning models are more widely studied than other methods, and that Grad-CAM, SHAP, LIME, and attention mechanisms are the most common explainability methods. The dataset of the Alzheimer's Disease Neuroimaging Initiative (ADNI) is widely used in existing studies, which is why Alzheimer's disease is still the major focus of study. While diagnostic capability is excellent, the translation to clinical practice is hampered by limited external validation, retrospective data sets, small sample sizes and regulatory issues. Future studies should focus on multimodal learning, prospective multi-center and various datasets along with trustworthy XAI frameworks to enable the routine clinical implementation.