MediKoGPT-Triage: A Korean Clinical Symptom Triage Dataset with ICD-10 Differential Diagnoses and Structured Clinical Reasoning (6,912 cases)

Overview MediKoGPT-Triage is a Korean-language clinical symptom triage dataset comprising 6,912 cases designed for training and evaluating healthcare generative AI systems. Each case provides a patient's chief complaint paired with structured clinical reasoning including ICD-10-coded differential diagnoses, severity-stratified triage flags, clinical assessment, and evidence-based recommendations. 본 데이터셋은 헬스케어 생성 AI 시스템의 학습 및 평가를 위해 구축된 6,912건 규모의 한국어 임상 증상 트리아지 데이터셋입니다. 각 사례는 환자의 주호소와 함께 ICD-10 코드 기반 감별진단, 중증도 트리아지 플래그, 임상 평가, 근거 기반 권장사항을 포함한 구조화된 임상 추론을 제공합니다. Dataset Composition Total cases: 6,912 Language: Korean, with ICD-10 standard codes Format: JSON Lines License: CC BY-NC 4.0 Data Schema Each record contains the following fields: Field Description data_id Unique case identifier flag Triage severity flag (red / yellow / green) title Chief complaint summary differential_diagnosis Array of differential diagnoses with ICD-10 codes analysis.summary Key clinical observations analysis.assessment Clinical assessment in natural language analysis.positive_aspects Positive clinical findings analysis.areas_of_concern Areas requiring clinical attention analysis.recommendations Categorized recommendations (medical / lifestyle) suggested_questions Follow-up questions for further history-taking medical_attention_needed Boolean flag for required medical attention medical_attention_reason Justification for medical attention recommendation disclaimer Optional disclaimer text Standards Compliance Differential diagnoses coded using ICD-10 (WHO, 10th Revision) Intended Use Training and evaluating Korean medical LLMs and triage agents Benchmarking structured clinical output quality Out-of-Scope This dataset is intended for RESEARCH USE ONLY and MUST NOT be used as a substitute for professional medical advice, diagnosis, or treatment. Outputs derived from models trained on this dataset should not be deployed in clinical settings without appropriate regulatory review and clinical validation. Acknowledgement This work was supported by the Starting growth Technological R&D Program (TIPS Program, RS-2024-00511590) funded by the Ministry of SMEs and Startups (Mss, Korea) in 2024