Neurodegenerative diseases like Alzheimer’s, Parkinson’s, vascular dementia, and frontotemporal dementia are difficult to detect early due to noisy or missing Mini-Mental State Examination (MMSE) data and uninterpretable prediction models. A Disentangled Conditional Variational Autoencoder (D-CVAE) Fusion model with a novel stage-aligned disentanglement loss addresses these issues. Clinically interpretable representations are achieved by separating cognitive decline components from other latent variations. The model reconstructs lost MMSE data and improves feature representation by conditioning the encoder-decoder architecture on cognitive stage. Using disentangled latent embeddings, an integrated Multi-Layer Perceptron (MLP) classifier accurately identifies cognitive impairment levels from no impairment to moderate impairment. Our D-CVAE Fusion model has 94.52% accuracy, resilience with 30% missing data, and clinically significant latent dimensions (e.g. z₁ shows strong correlation with MMSE score, r = − 0.94). Clinical decision support systems for cognitive evaluation and early-stage neurodegenerative disease diagnosis can use the suggested paradigm with confidence and interpretability.
Yarushev et al. (Mon,) studied this question.