EHRSHOT

EHRSHOT是由斯坦福大学创建的一个电子健康记录（EHR）数据集，包含6,739名患者的去标识化结构化数据。该数据集不同于MIMIC-III/IV等流行数据集，它涵盖了非ICU/ED患者的纵向健康数据。数据集内容包括人口统计、诊断、程序、实验室结果、药物和其他结构化数据，总计4160万次临床事件和921,499次访问。创建过程涉及从367万名患者的数据库中筛选出特定患者群体，并定义了15个少样本临床预测任务。EHRSHOT的应用领域包括评估基础模型在样本效率和任务适应性方面的表现，旨在解决医疗领域中机器学习模型的可重复性和数据访问问题。

EHRSHOT is an electronic health record (EHR) dataset created by Stanford University, containing de-identified structured data from 6,739 patients. Unlike popular datasets such as MIMIC-III/IV, this dataset covers longitudinal health data of non-ICU/ED patients. The dataset includes demographic information, diagnoses, procedures, laboratory results, medications and other structured data, totaling 41.6 million clinical events and 921,499 patient visits. Its development involved screening a specific patient cohort from a database of 3.67 million patients, and defining 15 few-shot clinical prediction tasks. Applications of EHRSHOT include evaluating the performance of foundation models in terms of sample efficiency and task adaptation, aiming to address the issues of reproducibility and data access for machine learning models in the healthcare domain.