Neuropsychiatric symptoms (NPS) are increasingly recognized as important components of disease progression in Alzheimer's disease (AD), yet the multivariate relationship between NPS burden and structural neurodegeneration remains incompletely characterized. This study sought to link sMRI-derived regional neurodegenerative biomarkers with NPS using an explainable deep learning approach. The authors aimed to predict NPS from regional sMRI measures and to provide individualized, interpretable regional attributions of neurodegeneration that relate to NPS across the AD continuum.
Analyses used data from two well-characterized repositories: the Alzheimer's Disease Neuroimaging Initiative (ADNI) and the Knight Alzheimer Disease Research Center (Knight-ADRC). The combined experimental dataset included N = 1,756 participants drawn from these cohorts. The manuscript reports that access to the underlying imaging and clinical data is governed by the respective repositories and may require registration and data-use agreements. Derived results generated in the study are available from the corresponding authors upon reasonable request and subject to applicable human-subject and data-transfer requirements.
The authors developed a deep neural network, designated NPSNet, to predict NPS scores from regional sMRI neurodegenerative biomarkers. Multiple machine learning regression models were compared for NPS prediction, including gradient boosting and support vector regression, among others. Model comparisons evaluated held-out correlation between predicted and observed NPS scores using Spearman correlation (ρ) and associated p values.
To provide interpretable regional attributions, the authors integrated three feature-importance methods with NPSNet: SHapley Additive exPlanations (SHAP), Local Interpretable Model-agnostic Explanations (LIME), and Layer-wise Relevance Propagation (LRP). To establish a ground truth for benchmarking explainability, the team introduced predefined regional perturbations into semi-simulated data and tested whether NPSNetSHAP, NPSNetLIME, and NPSNetLRP could recover the perturbed regions.
In pooled semi-simulated data, the NPSNet produced strong predictive performance for NPS scores, with a pooled Spearman correlation of ρ = 0.927 (p < 2.2 × 10⁻³⁰⁸). Fold-wise correlations ranged from ρ = 0.912 to ρ = 0.963. When recovering predefined regional perturbations in this semi-simulated setting, NPSNetSHAP recovered all perturbed regions (100%), NPSNetLIME recovered 90% of perturbed regions, and NPSNetLRP recovered 40%.
In the held-out experimental dataset including N = 1,756 participants, NPSNet achieved the highest held-out Spearman correlation for NPS prediction, ρ = 0.387 (p = 1.7 × 10⁻¹⁵). This performance exceeded that of gradient boosting (ρ = 0.301), support vector regression (ρ = 0.266), and other compared methods (each ρ < 0.266). These comparisons indicate that the deep learning approach provided superior predictive accuracy for NPS from regional sMRI measures in this multi-cohort sample.
Using NPSNet combined with SHAP (NPSNetSHAP), the authors produced individual-level regional contribution patterns relevant to predicted NPS burden. Comparing SHAP attributions between cognitively normal (CN) participants and those with mild cognitive impairment (MCI) or AD revealed distributed multivariate neurodegenerative signatures associated with NPS. The largest differences in attribution patterns were observed between CN and AD participants, indicating progressively altered distributed regional contributions to NPS along the disease continuum.
Data used in this study were obtained from ADNI and Knight-ADRC and are subject to the repositories’ access and data-use procedures. The Institutional Review Board of Washington University in St. Louis provided ethical approval for the work. The authors declared no competing interests. The study reports that any necessary patient/participant consent and institutional approvals were obtained, and that derived results are available from the corresponding authors on reasonable request and subject to applicable restrictions.
This work introduces an explainable deep learning framework that links regional sMRI neurodegenerative biomarkers to neuropsychiatric symptoms across the Alzheimer's disease continuum. The deep neural network NPSNet outperformed several established machine learning regressors for NPS prediction in held-out data, and the integration of SHAP produced robust, individualized regional attributions that recovered semi-simulated perturbations and revealed distributed neurodegenerative signatures that differ across diagnostic groups. The framework provides a reproducible approach for mapping individualized, distributed brain contributions to NPS burden using existing multi-cohort imaging resources.