Conformal Memory-Augmented Attention Networks for Robust and Adaptive Disease Prediction

: We propose a novel inference framework for longitudinal disease prediction, replacing conventional static
classifiers with a conformal memory-augmented attention network. The system processes multi-modal clinical time series by means of a temporal convolutional encoder, then applies a tensorized attention mechanism that retrieves prototypical patient trajectories from a dynamic memory bank. Instead of producing point estimates, our method generates statistically rigorous prediction sets with guaranteed coverage probabilities through a distribution-free conformal calibration layer embedded
directly into the attention computation. The nonconformity score measures how well a patient’s temporal embedding aligns with class-specific memory prototypes, and the resulting prediction sets adapt to distribution shifts without requiring retraining or post-hoc recalibration. During inference, a continual learning module refreshes memory banks via a prototypical replay mechanism that applies frequency-weighted consolidation, thus retaining previously learned patterns while accommodating novel data. The tensorized bilinear interaction between queries and memory prototypes captures higher-order feature relationships via a low-rank factorization that reduces parameters and improves generalization. Our approach
consequently yields robust, interpretable predictions that stay valid under label shifts and changing clinical data distributions. The system outputs both the prediction set and the most influential memory prototypes, thereby delivering clinicians actionable insights alongside statistically guaranteed uncertainty quantification.