A Privacy-Preserving OCR-LLM System for Coronary Syndrome Subtyping From Admission HPI: Multicenter Validation in China and the US (OCR-LLM-CHD)

Development and Multicenter Validation of a Privacy-Preserving OCR-LLM Pipeline for Four-Subtype Coronary Syndrome Classification Using Admission HPI Across Heterogeneous EHR Systems

This study develops and validates a privacy-preserving OCR-LLM pipeline that converts admission history of present illness (HPI) records into structured coronary syndrome subtypes (STEMI, NSTEMI, unstable angina, and chronic coronary syndrome). The system first extracts text from de-identified HPI images using locally deployed OCR, then applies large language models with a fixed diagnostic prompt to generate subtype classification and evidence. Performance is evaluated in an internal validation cohort and multiple external datasets covering heterogeneous EHR templates, emergency department cases, and an English dataset from MIMIC-IV. A clinician usability study assesses changes in diagnostic accuracy and time with and without tool assistance.

Study Overview

• Status: Not yet recruiting
• Conditions: Acute Coronary Syndromes; ST-segment Elevation Myocardial Infarction (STEMI); Coronary Artery Disease (CAD) (E.G., Angina, Myocardial Infarction, and Atherosclerotic Heart Disease (ASHD)); Non-ST-Segment Elevation Myocardial Infarction (NSTEMI)
• Intervention / Treatment: Device: OCR-Prompt-LLM Information Extraction and Classification Workflow (OCR-Prompt-LLM); Device: Manual Clinical Data Review

• Study Type: Observational
• Enrollment (Estimated): 10

Contacts and Locations

• Study Contact: Name: Xiaojin Gao, Dr; Phone Number: +86010 88322415; Email: sophie_gao@sina.com

Participation Criteria

Eligibility Criteria

• Ages Eligible for Study: Adult; Older Adult
• Accepts Healthy Volunteers: N/A

• Sampling Method: Probability Sample

Study Population

The study population includes (1) de-identified clinical encounter records and (2) physician participants for usability testing. For record-based cohorts, eligible encounters contain an admission/presentation History of Present Illness (HPI) (text or image-derived text) sufficient for 4-class coronary syndrome subtyping (e.g., STEMI, NSTEMI, unstable angina, chronic coronary syndrome). Records are analyzed at the encounter level and are organized into five dataset-based cohorts (internal validation, multicenter external validation, ED external validation, English EHR external validation, and clinician usability). Reference labels are assigned using a prespecified clinical adjudication process.

Description

Inclusion Criteria:Hospital encounters with admission HPI documenting sym

2a4afaef-9dc1-47fc-874f-9dffaf7…

evant to coronary syndrome subtyping.

Cases with sufficient documentation to assign one of four target subtypes (STEMI, NSTEMI, UA, CCS) by adjudication.

-

Exclusion Criteria: Unclear subtype or incomplete/uncertain time information preventing gold standard assignment.

Non-CHD primary reason for admission after screening (for MIMIC-IV cohort).

-

Study Plan

How is the study designed?

Number of groups / cohorts

5

Cohorts and Interventions

Group / Cohort	Intervention / Treatment
Internal Development and Validation Cohort; Retrospective cohort used for model development and internal validation. Inputs are de-identified admission HPI records (image or text) from the AIM-CHD dataset. Expert adjudication provides the reference standard labels for 4-class coronary syndrome subtyping (STEMI, NSTEMI, unstable angina, chronic coronary syndrome).	Device: OCR-Prompt-LLM Information Extraction and Classification Workflow (OCR-Prompt-LLM); An automated clinical data management workflow integrating Optical Character Recognition (OCR), optimized prompt engineering, and large language models (LLMs). The system processes unstructured inpatient/ED records (primarily admission history of present illness and related narrative text) to extract prespecified key clinical indicators (e.g., left ventricular ejection fraction, coronary syndrome subtype, medications) and to classify cases into prespecified coronary artery disease categories (e.g., unstable angina, STEMI, NSTEMI, chronic coronary syndrome). The workflow outputs structured fields and a classification result with supporting evidence excerpts.; Device: Manual Clinical Data Review; Standard manual process in which experienced clinicians review patient medical records and extract the same prespecified clinical indicators and coronary artery disease categories using routine clinical judgment and documentation review. This manual abstraction serves as the human benchmark for comparing diagnostic accuracy, completeness, and operational efficiency against the automated OCR-Prompt-LLM workflow.
Multicenter External Validation Cohort; Retrospective multicenter cohort used for external validation across heterogeneous EHR templates and documentation styles. De-identified admission HPI records are processed through the same OCR-LLM pipeline, and predictions are compared with expert adjudicated reference labels to assess generalizability.	Device: OCR-Prompt-LLM Information Extraction and Classification Workflow (OCR-Prompt-LLM); An automated clinical data management workflow integrating Optical Character Recognition (OCR), optimized prompt engineering, and large language models (LLMs). The system processes unstructured inpatient/ED records (primarily admission history of present illness and related narrative text) to extract prespecified key clinical indicators (e.g., left ventricular ejection fraction, coronary syndrome subtype, medications) and to classify cases into prespecified coronary artery disease categories (e.g., unstable angina, STEMI, NSTEMI, chronic coronary syndrome). The workflow outputs structured fields and a classification result with supporting evidence excerpts.; Device: Manual Clinical Data Review; Standard manual process in which experienced clinicians review patient medical records and extract the same prespecified clinical indicators and coronary artery disease categories using routine clinical judgment and documentation review. This manual abstraction serves as the human benchmark for comparing diagnostic accuracy, completeness, and operational efficiency against the automated OCR-Prompt-LLM workflow.
Emergency Department External Validation Cohort; Retrospective cohort representing real-world emergency department workflow. De-identified ED admission HPI records are used to evaluate model performance under time-sensitive, information-limited conditions and assess robustness to ED documentation variability.	Device: OCR-Prompt-LLM Information Extraction and Classification Workflow (OCR-Prompt-LLM); An automated clinical data management workflow integrating Optical Character Recognition (OCR), optimized prompt engineering, and large language models (LLMs). The system processes unstructured inpatient/ED records (primarily admission history of present illness and related narrative text) to extract prespecified key clinical indicators (e.g., left ventricular ejection fraction, coronary syndrome subtype, medications) and to classify cases into prespecified coronary artery disease categories (e.g., unstable angina, STEMI, NSTEMI, chronic coronary syndrome). The workflow outputs structured fields and a classification result with supporting evidence excerpts.; Device: Manual Clinical Data Review; Standard manual process in which experienced clinicians review patient medical records and extract the same prespecified clinical indicators and coronary artery disease categories using routine clinical judgment and documentation review. This manual abstraction serves as the human benchmark for comparing diagnostic accuracy, completeness, and operational efficiency against the automated OCR-Prompt-LLM workflow.
English EHR External Validation Cohort; Retrospective cohort derived from the public de-identified MIMIC-IV database. English admission notes/HPI text are used to evaluate cross-language transportability and performance of the same classification prompts and post-processing rules against reference labels derived by adjudication/structured diagnosis mapping (as prespecified in the protocol).	Device: OCR-Prompt-LLM Information Extraction and Classification Workflow (OCR-Prompt-LLM); An automated clinical data management workflow integrating Optical Character Recognition (OCR), optimized prompt engineering, and large language models (LLMs). The system processes unstructured inpatient/ED records (primarily admission history of present illness and related narrative text) to extract prespecified key clinical indicators (e.g., left ventricular ejection fraction, coronary syndrome subtype, medications) and to classify cases into prespecified coronary artery disease categories (e.g., unstable angina, STEMI, NSTEMI, chronic coronary syndrome). The workflow outputs structured fields and a classification result with supporting evidence excerpts.; Device: Manual Clinical Data Review; Standard manual process in which experienced clinicians review patient medical records and extract the same prespecified clinical indicators and coronary artery disease categories using routine clinical judgment and documentation review. This manual abstraction serves as the human benchmark for comparing diagnostic accuracy, completeness, and operational efficiency against the automated OCR-Prompt-LLM workflow.
Clinician Usability Cohort; Prospective usability evaluation cohort. Physicians complete a structured coronary syndrome subtyping task using admission HPI cases. Outcomes include diagnostic accuracy and time to completion; within-participant comparisons may be performed between unassisted and tool-assisted conditions as prespecified.	Device: OCR-Prompt-LLM Information Extraction and Classification Workflow (OCR-Prompt-LLM); An automated clinical data management workflow integrating Optical Character Recognition (OCR), optimized prompt engineering, and large language models (LLMs). The system processes unstructured inpatient/ED records (primarily admission history of present illness and related narrative text) to extract prespecified key clinical indicators (e.g., left ventricular ejection fraction, coronary syndrome subtype, medications) and to classify cases into prespecified coronary artery disease categories (e.g., unstable angina, STEMI, NSTEMI, chronic coronary syndrome). The workflow outputs structured fields and a classification result with supporting evidence excerpts.; Device: Manual Clinical Data Review; Standard manual process in which experienced clinicians review patient medical records and extract the same prespecified clinical indicators and coronary artery disease categories using routine clinical judgment and documentation review. This manual abstraction serves as the human benchmark for comparing diagnostic accuracy, completeness, and operational efficiency against the automated OCR-Prompt-LLM workflow.

What is the study measuring?

Primary Outcome Measures

Outcome Measure	Measure Description	Time Frame
Overall classification accuracy	Time Frame: Up to completion of dataset evaluation (internal + external cohorts) Description: Proportion of cases with correct subtype (STEMI/NSTEMI/UA/CCS) compared with expert-adjudicated gold standard.	1 month

Collaborators and Investigators

• Sponsor: China National Center for Cardiovascular Diseases

Study record dates

Study Major Dates

• Study Start (Estimated): February 28, 2026
• Primary Completion (Estimated): March 8, 2026
• Study Completion (Estimated): March 8, 2026

Study Registration Dates

• First Submitted: February 27, 2026
• First Submitted That Met QC Criteria: February 27, 2026
• First Posted (Actual): March 4, 2026

Study Record Updates

• Last Update Posted (Actual): March 4, 2026
• Last Update Submitted That Met QC Criteria: February 27, 2026
• Last Verified: February 1, 2026

More Information

Terms related to this study

• Additional Relevant MeSH Terms: Pain; Neurologic Manifestations; Vascular Diseases; Cardiovascular Diseases; Pathologic Processes; Heart Diseases; Infarction; Necrosis; Arteriosclerosis; Arterial Occlusive Diseases; Coronary Disease; Myocardial Ischemia; Ischemia; Chest Pain; Pathological Conditions, Signs and Symptoms; Signs and Symptoms; ST Elevation Myocardial Infarction; Non-ST Elevated Myocardial Infarction; Coronary Artery Disease; Myocardial Infarction; Acute Coronary Syndrome; Angina Pectoris
• Other Study ID Numbers: CAD-LLM-002; Sponsor (Other Grant/Funding Number: China National Center for Cardiovascular Diseases)

Plan for Individual participant data (IPD)

• Plan to Share Individual Participant Data (IPD)?: NO
• IPD Plan Description: To protect patient privacy and comply with the data management policies of the participating institutions (Fuwai Hospital and sub-centers), individual participant data will not be made publicly available. However, aggregated study results and statistical analyses will be included in the final publication.

Drug and device information, study documents

• Studies a U.S. FDA-regulated drug product: No
• Studies a U.S. FDA-regulated device product: No