metadata
license: cc-by-nc-sa-4.0
tags:
- medical
extra_gated_prompt: >-
You confirm that you fully understand the following: 1. This dataset contains
de-identified psychiatric inpatient electronic medical records and may only be
used for non-commercial academic research or clinical education; 2. You will
not attempt to re-identify patients by linking this dataset with other
information, nor transfer/share the dataset with third parties; 3. You will
comply with the Personal Information Protection Law of the People's Republic
of China and international medical privacy standards, and assume legal
responsibility for any non-compliant use.
extra_gated_fields:
Full Name of Affiliated Institution: text
Institution Type:
type: select
options:
- Academic Institution (University/Research Institute)
- Hospital/Clinical Center (Psychiatry Department)
- Non-Profit Organization (Mental Health Focus)
- Government/Regulatory Body
- label: Other
value: other
Specific Usage Scenario:
type: select
options:
- >-
Psychiatric LLM Performance Evaluation (e.g., diagnosis/medication
recommendation tasks)
- Psychiatric Clinical Text Analysis Research
- Medical Education (e.g., case teaching materials)
- label: Other (Please supplement details in "Purpose Description")
value: other
Purpose Description (Supplement detailed research objectives if "Other" is selected): text
Principal Investigator Name: text
Principal Investigator Contact Information (Email/Phone): text
I have obtained the complete ethical approval document, and the intended use is within the approved scope: checkbox
I commit to using the dataset only for non-commercial purposes and will not distribute/resell it: checkbox
I commit to adopting measures such as encrypted storage and access control to protect data security: checkbox
PsychBench
💻Github Repo 🖨️[Paper Under Review]
The official dataset for "PsychBench: A comprehensive and professional benchmark for evaluating the performance of LLM-assisted psychiatric clinical practice".
Introduction
PsychBench’s dataset is a high-quality, real-world psychiatric clinical resource built to support LLM evaluation in psychiatric practice. It draws from 900 de-identified electronic medical records of inpatients (no synthetic augmentation or paraphrasing), ensuring authenticity and alignment with real clinical scenarios. The dataset adheres to the international ICD-10 standards (covering F00-F99), making it a reliable foundation for assessing LLM utility in psychiatry.
Distribution Composition
- Multiple Regions and Ethnic Groups
The 900 cases are curated from three geographically and culturally diverse psychiatric medical centers, ensuring broad representativeness:
- Beijing Anding Hospital, Capital Medical University (Northern China): National Medical Center for Mental Disorders (Grade-III-A), with over 800 beds and 860,000+ annual outpatient visits—covers severe and complex psychiatric cases.
- Fourth People’s Hospital of Wuhu (Central China): Regional referral & primary care center (1,760 beds, 150,000+ annual outpatients)—reflects both specialized and provincial-level psychiatric services.
- Second People’s Hospital of Dali (Southwest China): Grade-III-A hospital in a multi-ethnic autonomous prefecture (serving 3M+ people)—includes culturally informed care data, particularly for ethnic minority populations.
The diagnostic distribution at each center matches its real-world inpatient population, and the dataset includes cases from multiple ethnic minorities, capturing diverse cultural and health belief backgrounds.
- Disease Category Coverage
The 900-case dataset spans five major psychiatric disorder groups, covering key areas of psychiatric clinical practice:
- Mood/affective disorders,
- Schizophrenic/delusional/personality disorders,
- Neurotic/stress-related/organic/somatoform disorders ,
- Substance-induced disorders,
- Developmental/childhood-onset disorders
- Data Fields
Each of the 900 cases includes comprehensive inpatient clinical records, with essential information for LLM evaluation:
- Demographic details (age, gender,...)
- Clinical histories: History of present illness, past medical history, personal & family history
- Treatment records: Medication orders, physician ward round notes
- Assessment data: Physical examinations, psychiatric evaluations, auxiliary test results
Core Evaluation Tasks
The 900-case dataset supports five core psychiatric clinical tasks (paired with psychiatry-specific quantitative metrics) to assess LLM performance in key clinical scenarios:
- Clinical Text Understanding
- Principal Diagnosis
- Differential Analysis
- Medication Recommendation
- Long-term Course Management