# MediTab: Scaling Medical Tabular Data Predictors via Data Consolidation, Enrichment, and Refinement

**Authors:** Zifeng Wang, Chufan Gao, Cao Xiao, Jimeng Sun  
**arXiv:** [2305.12081](https://arxiv.org/abs/2305.12081)  
**Date:** 2023-05-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.12081v4/

## Key points

- Traditional methods in medical applications have focused on algorithm design without considering data engineering
- Tabular datasets in medical settings exhibit heterogeneity and limited sample sizes per source
- MediTab introduces a method to scale medical tabular data predictors by leveraging large language models (LLMs) for data consolidation
- The approach involves aligning out-domain data with the target task through a "learn, annotate, and refinement" pipeline
- Pre-trained MediTab models can effectively infer for arbitrary tabular input within the domain without requiring fine-tuning
- Achieved significant improvements over supervised baselines in patient outcome prediction and trial outcome prediction datasets
- Demonstrated impressive zero-shot performances by outperforming supervised XGBoost models in two prediction tasks

## Summary

, , , , 

MediTab: Scaling Medical Tabular Data Predictors via Data Consolidation, Enrichment, and Refinement

In the field of medical applications, tabular data prediction plays a crucial role in tasks such as patient health risk prediction. However, traditional methods have primarily focused on algorithm design without giving due consideration to the importance of data engineering. This is particularly problematic in medical settings where tabular datasets often exhibit significant heterogeneity across different sources and have limited sample sizes per source. As a result, predictors trained on manually curated small datasets struggle to generalize across various tabular datasets during inference. To address these challenges, this paper introduces MediTab, a method designed to scale medical tabular data predictors to accommodate various tabular inputs with varying features. The key innovation lies in the use of a data engine that leverages large language models (LLMs) to consolidate tabular samples from diverse sources and overcome barriers posed by tables with distinct schemas. Additionally, MediTab aligns out-domain data with the target task through a "learn, annotate, and refinement" pipeline. By expanding the training data using this approach, pre-trained MediTab models can effectively infer for arbitrary tabular input within the domain without requiring fine-tuning. This results in significant improvements over supervised baselines, achieving an average ranking of 1.57 on 7 patient outcome prediction datasets and 1.00 on 3 trial outcome prediction datasets. Furthermore, MediTab demonstrates impressive zero-shot performances by outperforming supervised XGBoost models by 8.9% and 17.2% on average in two prediction tasks. In conclusion, this study presents a novel approach to training universal tabular data predictors for medical applications that addresses challenges related to limited data availability, inconsistent dataset structures, and varying prediction targets across domains. By generating large-scale training data through a combination of in-domain and out-domain datasets, MediTab showcases superior performance without requiring extensive modifications or retraining.