LongMoE: Longitudinal Multimodal Learning via Trajectory-Aware Mixture-of-Experts

Researchers introduce LongMoE, a machine learning framework designed to improve clinical AI systems by simultaneously handling missing patient data and tracking disease progression over time. The model combines mixture-of-experts routing with temporal pattern recognition, demonstrating improvements across major medical datasets (ADNI, OASIS-3, MIMIC-IV).

LongMoE addresses a critical gap in clinical AI development by tackling two interconnected problems that existing approaches handle separately. Real-world medical data rarely arrives complete—patients skip appointments, imaging fails, or records go missing—yet current longitudinal models assume full modality availability while missing-data frameworks ignore temporal context entirely. This disconnect limits deployment in actual clinical settings where incomplete information is the norm rather than exception.

The framework's innovation lies in its architecture combining context-aware imputation with trajectory-aware encoding. By leveraging frequency-domain analysis of irregular visit sequences, LongMoE captures how disease patterns evolve distinctly for different patients, then routes information through specialized expert modules selected specifically for each patient's clinical context. This personalized expert selection prevents the model from applying one-size-fits-all logic to heterogeneous patient populations.

For healthcare AI development, this represents meaningful progress toward production-ready systems. Clinical institutions struggle with multimodal incompleteness daily—radiology departments malfunction, EHR integrations fail, or patients decline certain tests. Models that degrade gracefully under these realistic conditions are substantially more valuable than those requiring perfect data. The experimental validation on three major datasets (covering neurodegeneration, Alzheimer's progression, and intensive care) suggests broad applicability across disease domains.

The work signals growing maturity in clinical AI, moving beyond idealized research conditions toward systems designed for messy real-world deployment. Healthcare organizations and medical AI developers should monitor whether such trajectory-aware approaches become standard practice, as they fundamentally change how clinical decision-support systems handle uncertainty and temporal dynamics.

• →LongMoE simultaneously addresses modality missingness and longitudinal disease progression, two challenges previously tackled separately by existing methods.
• →The framework uses trajectory-aware mixture-of-experts routing to select patient-specific expert modules based on clinical context and disease evolution.
• →Performance testing on ADNI, OASIS-3, and MIMIC-IV datasets demonstrates robustness under incomplete modality conditions while remaining competitive with full-data settings.
• →The model processes irregular visit sequences via frequency-domain temporal pattern analysis, capturing how diagnostic significance changes as patient disease states evolve.
• →This approach addresses practical clinical deployment challenges where complete multimodal data is rarely available in real-world healthcare settings.