AdaPCLA: Adaptive Prior-Calibrated Logit Adjustment for Long-Tailed Longitudinal EHR Generation
AdaPCLA is a framework for generating rare events in Electronic Health Records (EHR) by adaptively adjusting model logits during training. It uses simulated annealing to internalize data knowledge parameters, ensuring rare codes are learned effectively. This allows the model to generate high-fidelity, tail-event data and adapt zero-shot to new patient populations without retraining.