Explainable Deep Learning Reveals Distributed Neurodegeneration Signatures of Neuropsychiatric Symptoms Across the Alzheimer's Continuum
Journal:
medRxiv
Published Date:
Aug 27, 2026
Abstract
Neuropsychiatric symptoms (NPS) are increasingly recognized as critical components of the disease progression in Alzheimer's disease (AD), yet their relationship with neurodegeneration remain poorly characterized. We investigated the multivariate relationships between structural MRI (sMRI)-based regional neurodegenerative biomarkers and NPS using the Alzheimer's Disease Neuroimaging Initiative (ADNI) and Knight Alzheimer Disease Research Center (Knight-ADRC) cohorts (N = 1,756). The machine learning regression models were compared for NPS prediction, and the best-performing deep neural network (named NPSNet) was integrated with three feature-importance methods: SHapley Additive exPlanations (SHAP), Local Interpretable Model-agnostic Explanations (LIME), and Layer-wise Relevance Propagation (LRP). To establish a known ground truth, we introduced predefined regional perturbations into semi-simulated data and tested whether NPSNetSHAP, NPSNetLIME, and NPSNetLRP could recover them. The NPSNet strongly predicted NPS scores (pooled Spearman {rho} = 0.927, p < 2.2 x 10-3; fold-wise {rho} = 0.912-0.963) and NPSNetSHAP recovered all 100% perturbed regions, compared with 90% for NPSNetLIME and 40% for NPSNetLRP. In the experimental data (N = 1,756), the NPSNet produced the highest held-out correlation ({rho} = 0.387, p = 1.7 x 10-{superscript 1}), exceeding gradient boosting ({rho} = 0.301), support vector regression ({rho} = 0.266), and others ({rho} < 0.266). NPSNetSHAP identified individual-level regional contribution patterns relevant to NPS predictions. Comparing NPSNetSHAP attributions between cognitively normal (CN) and mild cognitive impairment (MCI)/AD groups revealed distributed multivariate neurodegenerative signatures of NPS, with the largest differences between CN and AD participants. This study introduces an explainable deep learning framework for identifying distributed, individualized neurodegeneration signatures of NPS burden across the AD continuum.
