AdaPCLA: Adaptive Prior-Calibrated Logit Adjustment for Long-Tailed Longitudinal EHR Generation

2026-07-14

AI Summary

This work addresses the challenge that existing autoregressive models struggle to accurately capture the co-occurrence structure of tail events—such as rare diseases—in longitudinal electronic health record (EHR) generation, leading to insufficient fidelity for rare subpopulations. The authors propose AdaPCLA, a novel framework that uniquely integrates prior calibration with simulated annealing training, enabling adaptive logit adjustment to internalize data distribution knowledge and support zero-shot cross-population distribution control. Theoretical analysis elucidates the logit update mechanism for rare events and establishes bounds on prior signal preservation under neural tangent kernel (NTK) conditions linked to the annealing schedule. Experiments on MIMIC-III and MIMIC-IV demonstrate that AdaPCLA improves the TailPairSeen metric by 114.2% and 65.1% over HALO, respectively, and achieves a 3.5% higher zero-shot cross-population F1 score compared to GPT-style generation.

Abstract

Generative modeling of longitudinal Electronic Health Records is increasingly important for privacy-preserving research, yet standard autoregressive models tend to underrepresent the co-occurrence structure of tail events (i.e., diseases, symptoms), reducing the fidelity and faithfulness of generated data for rare subpopulations. To this end, we propose AdaPCLA framework, which enables generative models to adaptively fit and generate EHR data through a data distribution-aware training strategy; this is achieved by internalizing data knowledge parameters by simulated annealing training. It also supports training-free adaptation to a diverse clinical population for generation through zero-shot distribution control. Moreover, our theoretical analysis characterizes rare-code logit updates through the label-wise empirical NTK and derives a prior-internalization bound for how annealing speed and NTK conditioning affect retained prior signals. Experiments on real-world data show that AdaPCLA achieves consistent gains in tail plausibility, downstream utility, and zero-shot control; in particular, it improves TailPairSeen over HALO by 114.2% on MIMIC-III and 65.1% on MIMIC-IV, outperforms GPT-style generation by 3.5% F1 for zero-shot cross-population adaptation.

Problem

• long-tailed
• longitudinal EHR generation
• tail events
• generative modeling
• rare subpopulations

Innovation

• AdaPCLA
• long-tailed EHR generation
• simulated annealing training
• zero-shot distribution control
• empirical NTK