EveryQuery: Zero-Shot Clinical Prediction via Task-Conditioned Pretraining over Electronic Health Records

Payal Chandak, Gregory Kondas, Isaac Kohane, Matthew McDermott · 2026

Read EveryQuery: Zero-Shot Clinical Prediction via Task-Conditioned Pretraining over Electronic Health Records by Payal Chandak, Gregory Kondas, Isaac Kohane, Matthew McDermott (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-07900.txt.

Description

Chandak and colleagues note that autoregressive health-record models predict by simulating many patient futures, which costs heavy compute and hurts rare events most. They take only a patient history as input, so users cannot prompt them directly. EveryQuery changes the setup: tasks are written as structured queries, and the model takes the query as a prompt beside the history, estimating outcomes without simulation or finetuning. Across three datasets it beat a strong baseline, with its largest edge on rare outcomes and far less inference compute.

Catalog record

CREATORPayal Chandak, Gregory Kondas, Isaac Kohane, Matthew McDermott
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE41023
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.07900

Browse the full library · About the Tower of Babel