CoPoE: structured product-of-experts for missing-modality Alzheimer’s diagnosis

A coordinate-based multimodal fusion architecture that maps available clinical, imaging, and fluid biomarkers into an interpretable RPNS space without synthetic data imputation.

Primary source: arXiv:2610.11394: https://arxiv.org/abs/2610.11394

What’s new: Clinical workflows for neurodegenerative disorders rarely collect every diagnostic modality for every patient. Missing MRI scans, omitted cerebrospinal fluid (CSF) draws, or absent genomic sequencing routinely hinder automated multimodal diagnostic systems. Standard solutions either synthesize missing data using generative surrogates (which risks introducing hallucinated artifacts) or project available features into unconstrained, opaque latent vectors. In CoPoE: Multimodal Fusion via Decomposable Disease-Coordinate Product-of-Experts for Missing-Modality Alzheimer’s Diagnosis (arXiv:2610.11394, 8 Oct 2026), An and Jang present CoPoE.

CoPoE structures its latent space along four clinically established axes: genetic Risk ($R$), molecular Pathology ($P$), Neurodegeneration ($N$), and clinical Stage ($S$). Each observed diagnostic channel parameterizes a diagonal Gaussian expert over this shared RPNS space. A masked Product-of-Experts (PoE) mechanism fuses only the present modalities by multiplying their probability densities, naturally handling arbitrary missing inputs without synthetic imputation. On the Alzheimer’s Disease Neuroimaging Initiative (ADNI) benchmark, CoPoE achieved top performance across all 15 observed modality subsets while significantly reducing Expected Calibration Error (ECE) and Brier scores. Furthermore, PET-supervised probing confirmed that tau pathology representations remained intact in the $P$-block even when fluid biomarkers were withheld.

Why it matters: Point-of-care medical AI must function flexibly across variable clinic resource levels without requiring expensive secondary scans purely to satisfy model inputs. By grounding the fusion bottleneck in biological disease axes, CoPoE provides both calibrated risk estimates and interpretable intermediate representations that clinicians can inspect.

Caveats: The RPNS coordinate structure is custom-tailored to the pathophysiological staging of Alzheimer’s disease; adapting the framework to multi-organ diseases or acute trauma requires defining alternative biological axes. Validation was conducted on the curated ADNI cohort, which features standardized acquisition protocols that may not fully reflect messy real-world EHR records.