Deep learning models to detect neurodegenerative diseases across different modalities, datasets, and stages: A...

BackgroundNeurodegenerative diseases like Alzheimer's and Parkinson's pose significant diagnostic challenges because they are incurable. Their symptoms often overlap with those of normal ageing and involve progressive neuronal loss. This makes early prediction, proper diagnosis, and ongoing monitoring crucial at every stage of the disease. Deep learning has become a useful tool for handling different types of data. However, the performance of these models relies heavily on the availability and variety of data, making it important to characterize datasets.ObjectiveThis PRISMA-based systematic review looks at 111 studies from 2019 to April 2026, organized into three clinical stages of neurodegenerative diseases: early prediction, diagnosis, and progression.MethodsEach study is assessed across four dimensions: data type (neuroimaging or non-neuroimaging), datasets (open-access, self-created, or hybrid), deep learning model architecture, and the use of explainable AI.ResultsMRI is the only imaging modality that spans all three stages. Speech is the most common non-neuroimaging method for diagnosis, while gait is used only for diagnosis and progression. The ADNI and PPMI datasets are the most commonly used. Self-created datasets introduce new modalities such as ECG and iPSCs, which are not found in open-access datasets. CNNs are the leading models for neuroimaging tasks, while LSTMs and RNNs capture temporal patterns in longitudinal datasets. Survival models are exclusively used for ALS and LATE. Fewer than 6% of the studies used confirmed XAI. Only 2 of the 111 studies validated their results on an independent external cohort.ConclusionThis review is the only one that covers all three disease stages across multiple neurodegenerative diseases.