公开胸部X线数据集的人口统计学报告:减轻深度学习模型中性别和种族差异的机会
该研究于2021年2月1日检索了超过100张胸片的公开胸部X线数据集,共纳入23个数据集(范围105-371858张图像),评估其人口统计学变量(年龄、种族或民族、性别、保险状态)的报告情况。多数数据集报告了年龄和性别,但仅少数报告种族或民族和保险状态,且数据集常对某一性别代表性不足(更常见为女性)。作者建议数据集报告标准人口统计学变量并尽可能平衡人口统计学代表性,同时建议研究人员在使用这些数据集时注意平衡人口统计学标签和疾病标签。
要点
- 23个公开胸部X线数据集中,报告种族或民族的仅2个(8.7%),报告保险状态的仅1个(4.3%)。
- 13个可获取性别分布的数据集中,男性平均占55.2%(范围47.8%-69.7%),8个(61.5%)男性占比过高,5个(38.5%)女性占比过高。
- 作者建议数据集报告标准人口统计学变量并尽可能平衡代表性,研究人员使用时应平衡人口统计学标签与疾病标签。
核心结论
- 多数公开胸部X线数据集在图像层面报告了年龄和性别,但很少报告种族或民族和保险状态,且数据集常对某一性别代表性不足(更常见为女性)。
研究设计与样本
- 检索2021年2月1日可获取的、图像数超过100张的公开胸部X线数据集,检索范围包括Radiopaedia和Kaggle等多种存储库,共纳入23个数据集(范围105-371858张图像)。
诊断标准与关键数值
- 23个数据集中,19个(82.6%)以某种形式报告人口统计学变量,17个(73.9%)在图像层面报告。
- 报告年龄的数据集19个(82.6%),报告性别的18个(78.2%),报告种族或民族的2个(8.7%),报告保险状态的1个(4.3%)。
- 13个可获取性别分布的数据集中,男性平均占55.2%,范围47.8%-69.7%。
- 上述13个数据集中,8个(61.5%)男性占比过高,5个(38.5%)女性占比过高。
临床应用与阅片要点
- 作者建议数据集报告标准人口统计学变量,并尽可能平衡人口统计学代表性以减轻偏倚。
- 作者建议研究人员在使用这些数据集时,除疾病标签外还应注意平衡人口统计学标签,并开发能处理这些不平衡的训练方法。
局限性
- 原文未报告局限性。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
英文原文(摘要)· 点击展开
OBJECTIVE: Data sets with demographic imbalances can introduce bias in deep learning models and potentially amplify existing health disparities. We evaluated the reporting of demographics and potential biases in publicly available chest radiograph (CXR) data sets.
METHODS: We reviewed publicly available CXR data sets available on February 1, 2021, with >100 CXRs and performed a thorough search of various repositories, including Radiopaedia and Kaggle. For each data set, we recorded the total number of images and whether the data set reported demographic variables (age, race or ethnicity, sex, insurance status) in aggregate and on an image-level basis.
RESULTS: Twenty-three CXR data sets were included (range, 105-371,858 images). Most data sets reported demographics in some form (19 of 23; 82.6%) and on an image level (17 of 23; 73.9%). The majority reported age (19 of 23; 82.6%) and sex (18 of 23; 78.2%), but a minority reported race or ethnicity (2 of 23; 8.7%) and insurance status (1 of 23; 4.3%). Of the 13 data sets with sex distribution readily available, the average breakdown was 55.2% male subjects, ranging from 47.8% to 69.7% male representation. Of these, 8 (61.5%) overrepresented male subjects and 5 (38.5%) overrepresented female subjects.
DISCUSSION: Although most publicly available CXR data sets report age and sex on an image-basis level, few report race or ethnicity and insurance status. Furthermore, these data sets frequently underrepresent one of the sexes, more frequently the female sex. We recommend that data sets report standard demographic variables, and when possible, balance demographic representation to mitigate bias. Furthermore, for researchers using these data sets, we recommend that attention be paid to balancing demographic labels in addition to disease labels, as well as developing training methods that can account for these imbalances.
引用
- Yi PH, Kim TK, Siegel E, et al. Demographic Reporting in Publicly Available Chest Radiograph Data Sets: Opportunities for Mitigating Sex and Racial Disparities in Deep Learning Models. J Am Coll Radiol. 2022. PMID: 35033310. doi:10.1016/j.jacr.2021.08.018
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。