aai530-group6/ddxplus|医疗诊断数据集|合成患者数据集

---
language:
- en
license: cc-by-4.0
license_link: https://creativecommons.org/licenses/by/4.0/
tags:
- automatic-diagnosis
- automatic-symptom-detection
- differential-diagnosis
- synthetic-patients
- diseases
- health-care
pretty_name: DDXPlus
size_categories:
- 1K<n<10K
source_datasets:
- original
task_categories:
- tabular-classification
task_ids:
- multi-class-classification
paperswithcode_id: ddxplus
configs:
- config_name: default
data_files:
- split: train
path: "train.csv"
- split: test
path: "test.csv"
- split: validate
path: "validate.csv"

extra_gated_prompt: "By accessing this dataset, you agree to use it solely for research purposes and not for clinical decision-making."
extra_gated_fields:
Consent: checkbox
Purpose of use:
type: select
options:
- Research
- Educational
- label: Other
value: other

train-eval-index:
- config: default
task: medical-diagnosis
task_id: binary-classification
splits:
train_split: train
eval_split: validate
col_mapping:
AGE: AGE
SEX: SEX
PATHOLOGY: PATHOLOGY
EVIDENCES: EVIDENCES
INITIAL_EVIDENCE: INITIAL_EVIDENCE
DIFFERENTIAL_DIAGNOSIS: DIFFERENTIAL_DIAGNOSIS
metrics:
- type: accuracy
name: Accuracy
- type: f1
name: F1 Score
---

# Dataset Description

We are releasing under the CC-BY licence a new large-scale dataset for Automatic Symptom Detection (ASD) and Automatic Diagnosis (AD) systems in the medical domain. The dataset contains patients synthesized using a proprietary medical knowledge base and a commercial rule-based AD system. Patients in the dataset are characterized by their socio-demographic data, a pathology they are suffering from, a set of symptoms and antecedents related to this pathology, and a differential diagnosis. The symptoms and antecedents can be binary, categorical and multi-choice, with the potential of leading to more efficient and natural interactions between ASD/AD systems and patients. To the best of our knowledge, this is the first large-scale dataset that includes the differential diagnosis, and non-binary symptoms and antecedents.

**Note**: We use evidence as a general term to refer to a symptom or an antecedent.

This directory contains the following files:
- **release_evidences.json**: a JSON file describing all possible evidences considered in the dataset.
- **release_conditions.json**: a JSON file describing all pathologies considered in the dataset.
- **release_train_patients.zip**: a CSV file containing the patients of the training set.
- **release_validate_patients.zip**: a CSV file containing the patients of the validation set.
- **release_test_patients.zip**: a CSV file containing the patients of the test set.

## Evidence Description

Each evidence in the `release_evidences.json` file is described using the following entries:
- **name**: name of the evidence.
- **code_question**: a code allowing to identify which evidences are related. Evidences having the same `code_question` form a group of related symptoms. The value of the `code_question` refers to the evidence that need to be simulated/activated for the other members of the group to be eventually simulated.
- **question_fr**: the query, in French, associated to the evidence.
- **question_en**: the query, in English, associated to the evidence.
- **is_antecedent**: a flag indicating whether the evidence is an antecedent or a symptom.
- **data_type**: the type of evidence. We use `B` for binary, `C` for categorical, and `M` for multi-choice evidences.
- **default_value**: the default value of the evidence. If this value is used to characterize the evidence, then it is as if the evidence was not synthesized.
- **possible-values**: the possible values for the evidences. Only valid for categorical and multi-choice evidences.
- **value_meaning**: The meaning, in French and English, of each code that is part of the `possible-values` field. Only valid for categorical and multi-choice evidences.

## Pathology Description
The file `release_conditions.json` contains information about the pathologies that patients in the datasets may suffer from. Each pathology has the following attributes:
- **condition_name**: name of the pathology.
- **cond-name-fr**: name of the pathology in French.
- **cond-name-eng**: name of the pathology in English.
- **icd10-id**: ICD-10 code of the pathology.
- **severity**: the severity associated with the pathology. The lower the more severe.
- **symptoms**: data structure describing the set of symptoms characterizing the pathology. Each symptom is represented by its corresponding `name` entry in the `release_evidences.json` file.
- **antecedents**: data structure describing the set of antecedents characterizing the pathology. Each antecedent is represented by its corresponding `name` entry in the `release_evidences.json` file.

## Patient Description

Each patient in each of the 3 sets has the following attributes:
- **AGE**: the age of the synthesized patient.
- **SEX**: the sex of the synthesized patient.
- **PATHOLOGY**: name of the ground truth pathology (`condition_name` property in the `release_conditions.json` file) that the synthesized patient is suffering from.
- **EVIDENCES**: list of evidences experienced by the patient. An evidence can either be binary, categorical or multi-choice. A categorical or multi-choice evidence is represented in the format `[evidence-name]_@_[evidence-value]` where [`evidence-name`] is the name of the evidence (`name` entry in the `release_evidences.json` file) and [`evidence-value`] is a value from the `possible-values` entry. Note that for a multi-choice evidence, it is possible to have several `[evidence-name]_@_[evidence-value]` items in the evidence list, with each item being associated with a different evidence value. A binary evidence is represented as `[evidence-name]`.
- **INITIAL_EVIDENCE**: the evidence provided by the patient to kick-start an interaction with an ASD/AD system. This is useful during model evaluation for a fair comparison of ASD/AD systems as they will all begin an interaction with a given patient from the same starting point. The initial evidence is randomly selected from the binary evidences found in the evidence list mentioned above (i.e., `EVIDENCES`) and it is part of this list.
- **DIFFERENTIAL_DIAGNOSIS**: The ground truth differential diagnosis for the patient. It is represented as a list of pairs of the form `[[patho_1, proba_1], [patho_2, proba_2], ...]` where `patho_i` is the pathology name (`condition_name` entry in the `release_conditions.json` file) and `proba_i` is its related probability.

## Note:

We hope this dataset will encourage future works for ASD and AD systems that consider the differential diagnosis and the severity of pathologies. It is important to keep in mind that this dataset is formed of synthetic patients and is meant for research purposes. Given the assumptions made during the generation process of this dataset, we would like to emphasize that the dataset should not be used to train and deploy a model prior to performing rigorous evaluations of the model performance and verifying that the system has proper coverage and representation of the population that it will interact with.

It is important to understand that the level of specificity, sensitivity and confidence that a physician will seek when evaluating a patient will be influenced by the clinical setting. The dataset was built for acute care and biased toward high mortality and morbidity pathologies. Physicians will tend to consider negative evidences as equally important in such a clinical context in order to evaluate high acuity diseases.

In the creation of the DDXPlus dataset, a small subset of the diseases was chosen to establish a baseline. Medical professionals have to consider this very important point when reviewing the results of models trained with this dataset, as the differential is considerably smaller. A smaller differential means less potential evidences to collect. It is thus essential to understand this point when we look at the differential produced and the evidence collected by a model based on this dataset.

For more information, please check our [paper](https://arxiv.org/abs/2205.09148).

原始信息汇总

数据集概述

基本信息

• 语言: 英语
• 许可证: CC-BY 4.0
• 标签: 自动诊断, 自动症状检测, 鉴别诊断, 合成患者, 疾病, 医疗保健
• 数据集名称: DDXPlus
• 数据集大小: 1K<n<10K
• 数据来源: 原始数据
• 任务类别: 表格分类
• 任务ID: 多类分类

配置信息

• 配置名称: default
• 数据文件:
  • 训练集: train.csv
  • 测试集: test.csv
  • 验证集: validate.csv

额外要求

• 访问条件: 仅用于研究目的，不用于临床决策。
• 字段:
  • 同意: 勾选框
  • 使用目的: 选择框 (研究, 教育, 其他)

训练与评估

• 配置: default
• 任务: 医疗诊断
• 任务ID: 二元分类
• 分割:
  • 训练集: train
  • 验证集: validate
• 列映射:
  • AGE: 年龄
  • SEX: 性别
  • PATHOLOGY: 病理
  • EVIDENCES: 证据
  • INITIAL_EVIDENCE: 初始证据
  • DIFFERENTIAL_DIAGNOSIS: 鉴别诊断
• 评估指标:
  • 准确率: Accuracy
  • F1分数: F1 Score

数据集描述

• 内容: 包含使用专有医学知识库和商业规则基础的自动诊断系统合成的患者数据。患者特征包括社会人口统计数据、病理、相关症状和先兆以及鉴别诊断。
• 证据: 症状或先兆的通用术语。

文件描述

• release_evidences.json: 描述数据集中所有可能的证据。
• release_conditions.json: 描述数据集中所有病理。
• release_train_patients.zip: 包含训练集患者的CSV文件。
• release_validate_patients.zip: 包含验证集患者的CSV文件。
• release_test_patients.zip: 包含测试集患者的CSV文件。

证据描述

• 字段:
  • name: 证据名称
  • code_question: 相关证据的代码
  • question_fr: 法语查询
  • question_en: 英语查询
  • is_antecedent: 是否为先兆
  • data_type: 证据类型 (B: 二元, C: 分类, M: 多选)
  • default_value: 默认值
  • possible-values: 可能值
  • value_meaning: 值的含义

病理描述

• 字段:
  • condition_name: 病理名称
  • cond-name-fr: 法语病理名称
  • cond-name-eng: 英语病理名称
  • icd10-id: ICD-10代码
  • severity: 严重程度
  • symptoms: 症状
  • antecedents: 先兆

患者描述

• 字段:
  • AGE: 年龄
  • SEX: 性别
  • PATHOLOGY: 病理
  • EVIDENCES: 证据
  • INITIAL_EVIDENCE: 初始证据
  • DIFFERENTIAL_DIAGNOSIS: 鉴别诊断

注意事项

• 数据集包含合成患者，仅用于研究目的。
• 在使用数据集训练和部署模型之前，需要进行严格的模型性能评估和验证。
• 数据集偏向于高死亡率和发病率的病理。

AI搜集汇总

数据集介绍

构建方式

DDXPlus数据集的构建基于专有的医疗知识库和商业规则为基础的诊断系统，通过合成患者信息的方式生成。该数据集整合了患者的社会人口学数据、患病病理学、与病理学相关的症状和病史，以及鉴别诊断信息。这一构建过程旨在模拟真实医疗环境中的患者情况，为自动症状检测和自动诊断系统提供训练和评估的基础。

使用方法

使用DDXPlus数据集时，研究人员需遵守Creative Commons BY 4.0许可协议。数据集提供了训练集、验证集和测试集，分别以CSV格式存储在压缩文件中。用户需解压文件后，根据数据集中的证据描述、病理学描述和患者描述等文档进行数据读取和分析。数据集适用于自动症状检测和自动诊断系统的研发和评估，但需注意，该数据集为合成数据，仅供研究使用，不应直接用于临床决策。

背景与挑战

背景概述

DDXPlus数据集是在医学领域自动症状检测与自动诊断系统的研究背景下诞生的。该数据集由合成患者构成，这些患者是基于专有医学知识库和商业规则基础诊断系统生成的。每个患者具有人口统计信息、所患疾病、与疾病相关的症状和前兆，以及鉴别诊断。该数据集是首个包含鉴别诊断和非二元症状与前兆的大规模数据集，其发布旨在推动自动症状检测与自动诊断系统的研究进展，并已获得Creative Commons BY 4.0许可证的授权。DDXPlus数据集的创建，为医学诊断领域提供了一个全新的研究资源，对于提升相关系统的交互效率和自然性具有重大意义。

当前挑战

DDXPlus数据集在构建过程中面临的挑战包括：确保合成患者的真实性和代表性，同时涵盖广泛的症状和前兆；在数据集中平衡疾病严重性和死亡率；以及处理鉴别诊断中疾病范围的限制问题。此外，数据集在应用研究中面临的挑战包括：如何有效利用鉴别诊断信息进行多类别分类任务，以及如何在模型训练和评估中考虑到临床环境中的特异性、敏感性和置信度要求。这些挑战对于研究人员来说既是机遇也是考验，需要他们在研究和应用中不断探索和优化。

常用场景

经典使用场景

在医学领域中，DDXPlus数据集的典型应用场景在于构建自动症状检测（ASD）和自动诊断（AD）系统。该数据集通过模拟患者的社会人口数据、患病情况、相关症状和病史以及鉴别诊断，为研究者提供了丰富的信息资源，以训练模型从而实现从患者数据中自动识别症状并做出初步诊断。

解决学术问题

DDXPlus数据集解决了传统诊断系统中缺乏全面鉴别诊断信息的问题，它提供了包含非二元症状和病史的数据，这对于提升诊断系统的准确性和效率至关重要。此外，该数据集为研究提供了包含不同严重程度疾病的数据，有助于学术研究中对疾病严重性和症状之间关系的探索。

实际应用

在实际应用中，DDXPlus数据集可用于开发和评估面向急性护理场景的诊断系统，特别是在处理高死亡率和高发病率疾病时。该数据集有助于提高医疗系统对急性疾病的响应速度和准确性，从而改善患者的治疗效果和生存率。

数据集最近研究

最新研究方向

DDXPlus数据集为医学领域内的自动症状检测与自动诊断系统提供了大规模的训练资源，其特色在于包含了差异诊断及非二元症状与先兆。近期研究聚焦于利用该数据集提升诊断模型的准确性与特异性，特别是在处理急性病痛与高死亡率病症时。该数据集促使研究者们探索更深入的模型架构，以及更精确的症状与差异诊断之间的关系，以期提高模型的临床适用性，为未来的医疗决策提供有力支持。