Researchers have developed a new framework called AdaPCLA to improve the generation of longitudinal Electronic Health Records (EHRs). Standard autoregressive models struggle with rare events, but AdaPCLA uses a data distribution-aware training strategy to enhance the fidelity and faithfulness of generated data, particularly for rare subpopulations. Experiments show AdaPCLA significantly outperforms existing methods like HALO and GPT-style generation in tail plausibility, downstream utility, and zero-shot cross-population adaptation on datasets such as MIMIC-III and MIMIC-IV. AI
IMPACT Improves the generation of rare events in healthcare data, potentially aiding privacy-preserving research and clinical population studies.
RANK_REASON This is a research paper detailing a new framework for generative modeling.
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 3 sources. How we write summaries →
