@inproceedings{nawal-etal-2026-mimic,
title = "{MIMIC}-{CTRL}-2026: A Realistic Clinical Trial Dataset to Benchmark {LLM} performance",
author = "Nawal, Noshin and
Margaryan, Robert and
Abdalla, Moustafa and
Ezekowitz, Justin A. and
Abdalla, Mohamed",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1297/",
doi = "10.18653/2026.findings-emnlp.1297",
pages = "27428--27452",
ISBN = "979-8-89176-462-0",
abstract = "Patient recruitment remains a major barrier in clinical trials, as identifying eligible participants often requires manual screening of complex health records, contributing significantly to trial delays and increased costs. While Large Language Models (LLMs) offer potential to automate this process, existing public datasets are overly simplistic in their patient representation: they pair full criterion lists with single-paragraph synthetic summaries, failing to capture the longitudinal complexity of real clinical records. We introduce MIMIC-CTRL-2026, a dataset of 800 manually labeled patient-criterion pairs grounded in real-world longitudinal EHR data from MIMIC-IV, annotated by clinically trained experts. Each instance pairs one eligibility criterion with a full patient record spanning multiple discharge notes, radiology reports, lab values, and medication histories. While simple prompting achieves high performance ({\ensuremath{\kappa}} = 0.84) on existing benchmarks, it collapses ({\ensuremath{\kappa}} = 0.06) on our dataset, revealing a substantial benchmark-difficulty gap between simplified and real-world clinical evaluation. Lastly, we present a framework combining structured pre-processing with a clinician-validated expanded prompt that substantially outperforms naive baselines without any model fine-tuning, establishing a realistic benchmark for automated clinical trial screening on real EHR data. Our dataset and code are available at: https://doi.org/10.5281/zenodo.22103011"
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="nawal-etal-2026-mimic">
<titleInfo>
<title>MIMIC-CTRL-2026: A Realistic Clinical Trial Dataset to Benchmark LLM performance</title>
</titleInfo>
<name type="personal">
<namePart type="given">Noshin</namePart>
<namePart type="family">Nawal</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Robert</namePart>
<namePart type="family">Margaryan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Moustafa</namePart>
<namePart type="family">Abdalla</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Justin</namePart>
<namePart type="given">A</namePart>
<namePart type="family">Ezekowitz</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mohamed</namePart>
<namePart type="family">Abdalla</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Patient recruitment remains a major barrier in clinical trials, as identifying eligible participants often requires manual screening of complex health records, contributing significantly to trial delays and increased costs. While Large Language Models (LLMs) offer potential to automate this process, existing public datasets are overly simplistic in their patient representation: they pair full criterion lists with single-paragraph synthetic summaries, failing to capture the longitudinal complexity of real clinical records. We introduce MIMIC-CTRL-2026, a dataset of 800 manually labeled patient-criterion pairs grounded in real-world longitudinal EHR data from MIMIC-IV, annotated by clinically trained experts. Each instance pairs one eligibility criterion with a full patient record spanning multiple discharge notes, radiology reports, lab values, and medication histories. While simple prompting achieves high performance (\ensuremathąppa = 0.84) on existing benchmarks, it collapses (\ensuremathąppa = 0.06) on our dataset, revealing a substantial benchmark-difficulty gap between simplified and real-world clinical evaluation. Lastly, we present a framework combining structured pre-processing with a clinician-validated expanded prompt that substantially outperforms naive baselines without any model fine-tuning, establishing a realistic benchmark for automated clinical trial screening on real EHR data. Our dataset and code are available at: https://doi.org/10.5281/zenodo.22103011</abstract>
<identifier type="citekey">nawal-etal-2026-mimic</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.1297</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1297/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>27428</start>
<end>27452</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T MIMIC-CTRL-2026: A Realistic Clinical Trial Dataset to Benchmark LLM performance
%A Nawal, Noshin
%A Margaryan, Robert
%A Abdalla, Moustafa
%A Ezekowitz, Justin A.
%A Abdalla, Mohamed
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F nawal-etal-2026-mimic
%X Patient recruitment remains a major barrier in clinical trials, as identifying eligible participants often requires manual screening of complex health records, contributing significantly to trial delays and increased costs. While Large Language Models (LLMs) offer potential to automate this process, existing public datasets are overly simplistic in their patient representation: they pair full criterion lists with single-paragraph synthetic summaries, failing to capture the longitudinal complexity of real clinical records. We introduce MIMIC-CTRL-2026, a dataset of 800 manually labeled patient-criterion pairs grounded in real-world longitudinal EHR data from MIMIC-IV, annotated by clinically trained experts. Each instance pairs one eligibility criterion with a full patient record spanning multiple discharge notes, radiology reports, lab values, and medication histories. While simple prompting achieves high performance (\ensuremathąppa = 0.84) on existing benchmarks, it collapses (\ensuremathąppa = 0.06) on our dataset, revealing a substantial benchmark-difficulty gap between simplified and real-world clinical evaluation. Lastly, we present a framework combining structured pre-processing with a clinician-validated expanded prompt that substantially outperforms naive baselines without any model fine-tuning, establishing a realistic benchmark for automated clinical trial screening on real EHR data. Our dataset and code are available at: https://doi.org/10.5281/zenodo.22103011
%R 10.18653/2026.findings-emnlp.1297
%U https://aclanthology.org/2026.findings-emnlp.1297/
%U https://doi.org/10.18653/2026.findings-emnlp.1297
%P 27428-27452
Markdown (Informal)
[MIMIC-CTRL-2026: A Realistic Clinical Trial Dataset to Benchmark LLM performance](https://aclanthology.org/2026.findings-emnlp.1297/) (Nawal et al., Findings 2026)
ACL