Chinese Medical Triage Dataset Overview
- Chinese Medical Triage Dataset is a collection of Chinese-language resources focused on department recommendation and clinical routing based on patient symptoms and consultation texts.
- It encompasses diverse paradigms from full admission EHR trajectories and multi-round online consultations to large-scale text classification, each with distinct data provenance and label structures.
- The datasets feature rigorous annotation, normalization, and evaluation protocols that provide actionable insights for AI-assisted clinical workflow design.
Chinese medical triage datasets are Chinese-language resources designed to support department recommendation, clinical routing, and related decision support from patient-provided symptoms, consultation text, or other medically structured observations. Recent work describes several distinct dataset paradigms: the Chinese component of ClinicalMC, which embeds triage in a full admission-to-discharge trajectory derived from hospital electronic health records; a 3,360-case online consultation corpus for collaborative multi-agent triage under uncertainty; LCMDC, which frames triage as large-scale coarse-grained and fine-grained text classification over online medical dialogues; and, in a related but distinct direction, TCM-Tongue, a standardized image dataset for triage-like screening and symptom recognition rather than department routing (Hou et al., 2 Jun 2026, Cheng et al., 30 Jul 2025, Wang et al., 2024, Jin et al., 24 Jul 2025).
1. Corpus families and data provenance
The Chinese triage literature represented in these resources differs primarily by source domain. In ClinicalMC, the Chinese subset is built from MedEureka EHRs and is explicitly described as hospital electronic health records rather than exam-style question answering. The benchmark is designed to cover the full admission-to-discharge workflow, and triage is the first stage of every trajectory. After anonymization, completeness filtering, outcome filtering, and duplicate removal, the Chinese pool is reduced from 6,947 EHRs to 5,106, then 4,179, then 3,317 high-quality Chinese EHRs; from this set, the final benchmark contains 1,275 Chinese samples. The paper does not report a separate triage-only subset count, because triage corresponds to the admission stage of the full 1,275 Chinese cases (Hou et al., 2 Jun 2026).
The multi-agent dynamic matching work constructs a Chinese medical triage dataset from Ai Ai Yi Medical Network / iiyi.com, collected with the Scrapy web scraping framework. Its source is an online medical consultation platform rather than a single hospital. Each case is a doctor-patient interaction record with structured fields including patient demographics, chief complaint, medical history or history of present illness, and the ground-truth department assignment determined by licensed physicians. The resulting dataset contains 3,360 real-world cases spanning 9 primary departments and 62 secondary departments (Cheng et al., 30 Jul 2025).
LCMDC is built from the Chinese online medical consultation platform “Quick Doctor” (120ask.com). It provides a Coarse-grained Triage dataset with 439,630 samples and 14 categories, and a Fine-grained Diagnosis dataset with 199,600 samples and 120 categories. The same crawl also supports a Medical Consultation dataset, but the triage-relevant parts are the coarse-grained and fine-grained classification corpora derived from the platform’s departmental hierarchy (Wang et al., 2024).
| Resource | Source | Reported triage scope |
|---|---|---|
| ClinicalMC Chinese subset | MedEureka EHRs | 1,275 Chinese samples; triage is the first stage; 16 departments |
| Multi-agent dynamic matching dataset | iiyi.com | 3,360 cases; 9 primary departments; 62 secondary departments |
| LCMDC triage corpora | 120ask.com | 439,630 coarse-grained samples and 199,600 fine-grained samples |
These datasets therefore occupy different points on the realism–structure spectrum: hospital EHR trajectories, online doctor-patient consultations with iterative questioning, and very large dialogue-derived text-classification corpora.
2. Task formulations and label structures
Across the core triage datasets, triage is defined as department recommendation, not disease diagnosis. ClinicalMC states this explicitly through the formulation
where is the chief complaint, is basic information, is candidate departments, and is the predicted department. The input therefore consists of chief complaint, basic information, and a candidate department set, while the output is a single department label. The paper reports that the Chinese dataset covers 16 departments and gives examples of imbalance: the largest Chinese department is Surgery with 391 EHRs, while Chinese Medicine and Stomatology have 6 EHRs each (Hou et al., 2 Jun 2026).
The multi-agent dynamic matching dataset uses the department label as the target variable and states that the task is department recommendation rather than diagnosis in the strict disease sense. Its label ontology is explicitly hierarchical: 9 primary departments and 62 secondary departments. The system’s internal representation is centered on HPI normalization, so symptoms are not treated as a flat bag of words but as a standardized clinical narrative containing symptom onset time, symptom characteristics, symptom progression, treatment history, and relevant prior information from the conversation (Cheng et al., 30 Jul 2025).
LCMDC formalizes triage as text classification:
where the model predicts a probability distribution over department labels from patient consultation text. The paper distinguishes coarse-grained department routing from fine-grained disease/diagnosis routing. The coarse-grained label space has 14 primary department categories, while the fine-grained label space has 120 detailed categories (Wang et al., 2024).
A related but non-equivalent formulation appears in TCM-Tongue. That dataset is not a department-routing resource; it is a standardized tongue-image dataset for AI-assisted traditional Chinese medicine diagnosis with 20 pathological symptom categories and an average of 2.54 labels per image. Its relevance to triage is explicitly framed as triage-like screening, symptom recognition, and decision support, rather than direct department assignment (Jin et al., 24 Jul 2025).
3. Placement of triage within clinical workflow
ClinicalMC is distinctive because triage is not isolated. The benchmark is organized into four stages: Triage; First-course examination / diagnosis / treatment; Subsequent multi-course examination / assessment / treatment; and Final diagnosis. Triage occurs at admission, uses only chief complaint + basic information, and outputs department selection. The later stages integrate present history, past history, physical examination, auxiliary tests, evolving assessments, treatment adjustments, and finally the entire trajectory at discharge. The Chinese subset has an average of 3.42 clinical courses per patient, with a range of 2 to 10; the course-count distribution is 789 cases with 2 courses, 440 with 3, 34 with 4, 5 with 5, 2 with 6, 1 with 7, 0 with 8, 1 with 9, and 0 with 10. This concentration in short trajectories means that triage functions as the entry point to a comparatively concise longitudinal reasoning chain (Hou et al., 2 Jun 2026).
The same paper embeds this process in the SimHospital framework, which includes patient agent, examiner agent, and doctor agent. At the first stage, the patient agent presents the chief complaint and the doctor agent recommends the appropriate department. Triage is therefore operationalized as a routing decision within a simulated clinical environment rather than as a stand-alone classifier (Hou et al., 2 Jun 2026).
The multi-agent dynamic matching system also models triage as a staged interaction, but its workflow is organized around PatientAgent, RecipientAgent, InquirerAgent, and DepartmentAgent. The system simulates 4-round doctor-patient interactions: the patient agent produces responses from case data, the recipient agent standardizes unstructured inputs into HPI, the inquirer agent asks targeted follow-up questions, and the department agent recommends the department. Here, the dataset is used to test whether iterative inquiry improves department recommendation under uncertainty (Cheng et al., 30 Jul 2025).
LCMDC takes a different approach. Its triage datasets are single-turn text classification resources, while the broader corpus also supports a separate medical consultation task in which the input is a patient question and the output is a doctor response. This separation makes LCMDC especially suitable for comparing routing-oriented classification and consultation-oriented generation within the same broad data ecosystem (Wang et al., 2024).
4. Annotation, normalization, and quality control
ClinicalMC reports a relatively elaborate annotation and review process for the Chinese subset. The appendix describes a three-round annotation workflow. In the first round, GPT-4o converts summaries into full EHR-like trajectories, generates initial treatment plans, and segments records into progress notes in SOAP format, with prompts that explicitly instruct the model not to hallucinate absent information. In the second round, three clinically trained medical students review the outputs; invalid samples are discarded or re-annotated, with validation criteria including course-count consistency, completeness and accuracy of examination information, and correct field semantics. In the third round, two clinicians perform dual expert review, 30% of cases are sampled for quality assessment, and the process repeats until sampling accuracy reaches 95% or higher. The paper further states that three senior clinicians independently review sampled cases for final dataset quality checking, and that for the Chinese subset 1,000 cases are sampled in an additional quality-control procedure whose six binary dimensions include accuracy of triage (Hou et al., 2 Jun 2026).
The multi-agent dynamic matching dataset emphasizes preprocessing and normalization rather than multi-stage expert adjudication. Missing or invalid fields such as "None" or "No Information" are handled through a three-part imputation pipeline using Alibaba Cloud’s Qwen-plus API: Integrity Detection, Prompt Construction, and LLM-based Completion. The case context is truncated to 1,000 characters for API compatibility, and the completion is constrained through explicit missing-field specification and JSON formatting constraints. The dataset also normalizes departmental structure into primary departments and secondary departments, with the stated constraint that primary and secondary departments cannot share the same name. In addition, the paper introduces department-specific guidance rules and department comparison rules that function as a task-specific decision ontology (Cheng et al., 30 Jul 2025).
ClinicalMC’s initial raw-data processing is also unusually explicit. Personal identifiers are detected with regular expressions and replaced with placeholders such as “Patient A” or random values; records lacking key clinical information such as chief complaints, diagnoses, or treatment processes are removed; cases whose final outcome is death are excluded; and duplicates are removed through fine-grained demographic matching such as gender and occupation. These choices define the benchmark’s admissible clinical trajectories and directly shape the Chinese triage component (Hou et al., 2 Jun 2026).
5. Evaluation protocols and reported performance
ClinicalMC uses Accuracy (Acc) as the explicit triage metric. Evaluation is conducted in two settings. In the single-turn static setting, downstream tasks receive ground-truth annotations from prior tasks, which isolates stage-specific performance from upstream errors. In the multi-turn dynamic setting, the model’s own earlier outputs are fed into later stages; triage, as the first stage, is unaffected by previous-stage error propagation but becomes the gateway decision affecting all subsequent stages. For Chinese triage in the static setting, selected accuracies include Apollo2-7B: 52.71, HuatuoGPT2-34B: 57.59, MedGemma: 65.65, HuatuoGPT-o1-7B: 65.73, Qwen2.5-32B: 59.92, GPT-5-mini: 59.62, and Qwen3-Next-80B-A3B: 68.31, which is the best triage score in the main Chinese static table. In the dynamic setting, reported Chinese triage results include HuatuoGPT2-34B: 44.86, Qwen2.5-7B: 59.81, and GPT-4o-mini: 63.55. The paper notes that Chinese dynamic performance can improve slightly relative to static evaluation for some models and attributes this to shorter Chinese trajectories and more concise information chains. It also reports a Human (sampling) average on Chinese of 85.00, while specifying that this human score is an average across tasks rather than triage alone (Hou et al., 2 Jun 2026).
The multi-agent dynamic matching work evaluates performance after four interaction rounds over all 3,360 cases. It reports 74.2% overall accuracy, 89.2% primary department accuracy, and 73.9% secondary department accuracy. Round-by-round results improve from 85.5% → 89.2% for primary departments and 66.6% → 73.9% for secondary departments. The paper also reports an average 3.65/5 clinical capability score across six dimensions, including Triage Accuracy: 4.25 and Diagnostic Reasoning: 3.26. Error distribution is given as 74.2% correct, 15.4% secondary-department errors, and 10.4% primary-department errors. Per-department examples include Internal Medicine: 83.2%, Obstetrics & Gynecology: 76.3%, Pediatrics: 30.0%, and Oncology: 35.4%. On a special subset of 258 “challenging yet solvable” cases, the Baseline model reaches 100% overall accuracy by round 4, while the No HPI model remains lower with a +33.1% gap versus baseline; ablations on guidance strategies report Dynamic Only: 62.0%, Comparison Only: 57.8%, and No Guidance: 45.3% overall accuracy on the same subset (Cheng et al., 30 Jul 2025).
LCMDC evaluates triage with Accuracy, Precision, Recall, and F1. On the Coarse-grained Triage dataset, the proposed method reports Accuracy: 81.34%, F1: 82.05%, Precision: 65.60%, and Recall: 66.54%; on the Fine-grained Diagnosis dataset, it reports Accuracy: 83.31%, F1: 74.12%, Precision: 77.31%, and Recall: 72.30%. The paper additionally tests prompt learning on smaller imbalanced subsets of 73,604 entries, 9 categories for coarse-grained triage and 24,940 entries, 83 categories for fine-grained triage, concluding that prompt learning is particularly effective when data are sparse and labels are rare (Wang et al., 2024).
| Resource | Evaluation setting | Reported headline result |
|---|---|---|
| ClinicalMC Chinese triage | Static Accuracy | Best reported triage score: 68.31 |
| Multi-agent dynamic matching | Four interaction rounds | 89.2% primary / 73.9% secondary / 74.2% overall |
| LCMDC | Proposed triage method | 81.34% coarse-grained accuracy; 83.31% fine-grained accuracy |
Taken together, these results show that Chinese medical triage performance depends strongly on corpus design: longitudinal EHR routing in ClinicalMC remains materially below the reported human average across tasks, multi-round questioning substantially improves hierarchical online triage, and very large dialogue corpora support stronger supervised text classification.
6. Limitations, boundary conditions, and related directions
The datasets also expose recurring limitations. In ClinicalMC, triage is embedded within full trajectories rather than released as a separate triage-only split; the dataset is text-only, lacking multimodal data such as imaging or physiological waveforms; the Chinese department distribution is imbalanced; and the paper provides limited triage-specific error analysis, with most error analysis focused on later-stage issues such as redundant diagnostic or treatment plans, failure to detect subtle but critical changes, incorrect clinical diagnosis, incorrect reasoning chains, and other errors (Hou et al., 2 Jun 2026).
The multi-agent dynamic matching dataset is explicitly anonymized and is positioned as more diverse than single-institution data, but the paper also indicates that class imbalance and varying difficulty remain important. It notes stronger performance in Internal Medicine and Obstetrics & Gynecology, weaker performance in Pediatrics and Oncology, and concentrated cross-department confusions, especially between Internal Medicine and Surgery. The authors also emphasize that ambiguous symptom descriptions and incomplete records are common in real-world consultation data. This is a source of realism, but it also means that online consultation data may differ from in-hospital emergency triage (Cheng et al., 30 Jul 2025).
LCMDC is unusually explicit about limitations: lack of emergency/severe cases, age imbalance, anonymity effects, no comparison with human doctors for triage reliability, no multimodal inputs such as images or audio, single-turn consultation only, and no role-playing / simulation realism in generated conversations. The demographic profile of the collected corpus is also skewed: Female users account for 54.99%, Male users for 41.00%, and Others for 4.01%, while the majority of users are in the 10–40 age range and older groups are underrepresented (Wang et al., 2024).
A related but separate direction is represented by TCM-Tongue, which should not be conflated with department-routing triage datasets. It contains 6,719 high-quality images collected under standardized conditions, annotated with 20 pathological symptom categories, and supports multiple annotation formats including COCO, TXT, and XML. Its intended use includes automated tongue diagnosis, multi-label symptom classification, and triage-like screening before full clinical evaluation, especially in contexts such as telemedicine or AI-assisted traditional Chinese medicine workflows. This suggests a broader Chinese triage-data landscape in which department recommendation corpora and standardized screening datasets coexist but address different operational definitions of triage (Jin et al., 24 Jul 2025).
The central distinction across all of these resources is therefore methodological rather than merely linguistic. ClinicalMC models triage as the first decision in a longitudinal hospital trajectory; the multi-agent dynamic matching dataset models triage as an interactive, uncertainty-reducing dialogue process over hierarchical departments; LCMDC models triage as large-scale text classification over online consultations; and TCM-Tongue models triage-adjacent screening through standardized medical imaging. Together, they define the present contours of Chinese medical triage data for research on department routing, interactive clinical questioning, longitudinal decision-making, and multimodal screening.