ReClaim

This repository contains code for ReClaim, a family of generative transformer models for longitudinal medical claims trajectories.

ReClaim is described in the paper Foundation Models to Unlock Real-World Evidence from Nationwide Medical Claims. The paper studies administrative claims as a substrate for healthcare foundation models, training ReClaim on nationwide MarketScan claims trajectories and evaluating it across disease-onset prediction, expenditure forecasting, and real-world evidence analyses.

Codebase Overview

The repository is organized as a research pipeline:

Path	Purpose
preprocessing/	Build vocabulary mappings and convert MarketScan, EHRSHOT, and YNHH data into ReClaim-compatible token sequences.
training/	Build Hugging Face tokenizers, prepare pre-training and post-training datasets, and train Qwen-style causal language models.
downstream_tasks/disease_onset_prediction/	Prepare disease case/control cohorts, run model inference, and compute disease-onset AUCs.
downstream_tasks/expenditure_forecasting/	Predict next-year healthcare expenditure from tokenized patient history using ReClaim generations and tabular baselines.
downstream_tasks/rwe_support/	Generate patient embeddings and evaluate their use in real-world evidence target-trial analyses.
figures/	Figure assets and editable slide materials.

Workflow

At a high level, the project follows this order:

1. Generate medical vocabulary mappings under preprocessing/vocab_mapping/.
2. Preprocess raw claims or OMOP-style EHR data into monthly token sequences.
3. Build a tokenizer and train or post-train ReClaim models under training/.
4. Run downstream evaluations for disease onset, cost forecasting, and RWE support.

Each major folder has its own README with the expected inputs, local path configuration, and stage-by-stage commands.

Notes

Large raw datasets, generated mappings, model checkpoints, logs, and most intermediate outputs are not included in this repository. Many scripts are designed for a Slurm/HPC environment and require local paths or environment variables to be configured before running.

Citation

If you use this code, please cite the ReClaim paper:

@misc{ma2026reclaim,
  title = {Foundation Models to Unlock Real-World Evidence from Nationwide Medical Claims},
  author = {Ma, Fan and Liu, Yuntian and Lan, Xiang and others},
  year = {2026},
  eprint = {2605.02740},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI}
}