[Paper Review] Multimodal normative modeling in Alzheimers Disease with introspective variational autoencoders

Sayantan Kumar, Peijie Qiu+1 more|arXiv (Cornell University)|Feb 8, 2026

Dementia and Cognitive Impairment Research0 citations

TL;DR

The paper introduces mmSIVAE, a multimodal soft-introspective VAE with MOPOE aggregation to improve healthy-reference fidelity and multimodal fusion for normative modeling in AD, yielding more discriminative deviation scores and interpretable regional abnormalities.

ABSTRACT

Normative modeling learns a healthy reference distribution and quantifies subject-specific deviations to capture heterogeneous disease effects. In Alzheimers disease (AD), multimodal neuroimaging offers complementary signals but VAE-based normative models often (i) fit the healthy reference distribution imperfectly, inflating false positives, and (ii) use posterior aggregation (e.g., PoE/MoE) that can yield weak multimodal fusion in the shared latent space. We propose mmSIVAE, a multimodal soft-introspective variational autoencoder combined with Mixture-of-Product-of-Experts (MOPOE) aggregation to improve reference fidelity and multimodal integration. We compute deviation scores in latent space and feature space as distances from the learned healthy distributions, and map statistically significant latent deviations to regional abnormalities for interpretability. On ADNI MRI regional volumes and amyloid PET SUVR, mmSIVAE improves reconstruction on held-out controls and produces more discriminative deviation scores for outlier detection than VAE baselines, with higher likelihood ratios and clearer separation between control and AD-spectrum cohorts. Deviation maps highlight region-level patterns aligned with established AD-related changes. More broadly, our results highlight the importance of training objectives that prioritize reference-distribution fidelity and robust multimodal posterior aggregation for normative modeling, with implications for deviation-based analysis across multimodal clinical data.

Motivation & Objective

• Quantify subject-specific deviations from a healthy reference distribution in Alzheimer's disease using multimodal neuroimaging.
• Improve fidelity of the healthy reference and robustness of multimodal posterior aggregation.
• Provide interpretable deviation maps linking latent deviations to regional brain abnormalities.
• Demonstrate improved outlier detection and clinically meaningful patterns across AD stages.
• Offer a general framework for deviation-based analysis in multimodal clinical data.

Proposed method

• Extend Soft-IntroVAE to multimodal data (mmSIVAE) with an introspective training objective.
• Adopt MOPOE (Mixture-of-Product-of-Experts) to robustly aggregate modality-specific posteriors into a shared latent space.
• Define multimodal normative deviation scores in the learned latent space and in feature space using Mahalanobis distance.
• Map significant latent deviations to regional feature-space deviations for interpretability.
• Provide theory on encoder–decoder game and Nash equilibrium for multimodal mmSIVAE.
• Evaluate reconstruction quality, outlier detection (likelihood ratios), and regional deviation maps on ADNI MRI volumes and amyloid PET SUVR.

Experimental results

Research questions

• RQ1 Does mmSIVAE improve reconstruction fidelity for healthy controls compared with baselines?
• RQ2 Do MOPOE-based latent space aggregations yield sharper, more informative joint latent representations than PoE or MoE alone?
• RQ3 Are subject-wise latent and feature-space deviation scores more discriminative for distinguishing AD-spectrum from controls?
• RQ4 Can latent deviations be mapped to region-level abnormalities aligned with AD pathology?
• RQ5 How do the proposed deviations vary across AD stages and cognitive measures?

Key findings

• mmSIVAE with MOPOE achieves higher discrimination in outlier detection (higher likelihood ratios) than baselines across latent dimensions.
• Latent-space deviations (D_ml) generally outperform feature-space deviations (D_mf) for detecting outliers in AD subjects.
• MOPOE aggregation yields stronger separation between control and AD cohorts than PoE or MOE alone, especially at higher latent dimensionalities (d up to 20).
• Reconstruction errors for mmSIVAE are lower than baselines for both MRI volumes and amyloid SUVR, indicating better healthy-reference fidelity.
• Deviations mapped to brain regions show patterns consistent with established AD pathology (temporal, parietal, hippocampal atrophy; accentuated amyloid uptake in predefined regions).
• Latent dimensions with significant deviations correspond to identifiable regional deviations, enabling interpretable deviation maps across Desikan–Killiany and Aseg atlases.