HARC: Missingness-Aware Multimodal Residual Correction for Longitudinal ADAS-13 Score Prediction in Alzheimer's Disease

Background and Objective: Alzheimer’s disease (AD) is characterized by nonlinear cognitive decline and substantial inter-individual heterogeneity. Longitudinal prediction of 13-item Alzheimer’s Disease Assessment Scale–Cognitive Subscale (ADAS-13) scores remains challenging because of irregular followup intervals, multimodal missingness, sparse imaging observations, and increasing uncertainty over long forecasting horizons. Existing multimodal approaches often rely on direct fusion and may therefore be vulnerable to incomplete modalities and sparse imaging evidence. We developed HARC, a horizon-aware residual correction framework that preserves a strong cognitivetrajectory backbone while using multimodal biomarkers to refine long-term predictions.Methods: HARC first uses an ADAS-only GRU-D backbone to model irregular longitudinal ADAS-13 trajectories. A residual correction module then incorporates cerebrospinal fluid biomarkers and sparse MRI, FDG-PET, and amyloid-PET features through missingness-aware gating, historical image memory, modality-specific residual pathways, and horizon-aware calibration.Results: Experiments were conducted on longitudinal data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), using subject-level data splitting and three training seeds. Compared with the ADAS-only backbone, HARC reduced MAE from 4.6687 ± 0.0478 to 4.4994 ± 0.0983 and RMSE from 7.0616 ± 0.0790 to 6.7321 ± 0.0226, and improved R² from 0.7315 ± 0.0060 to 0.7560 ± 0.0016. Improvements were more pronounced at medium and long forecasting horizons, while short-term performance remained stable.Conclusions: HARC provides a practical and interpretable strategy for longitudinal ADAS-13 prediction under incomplete multimodal observations by using pathological and imaging evidence as conservative residual corrections rather than replacing cognitive-trajectory modeling.