Health Record Hiccups - 5526 real-world time series with change points labelled by crowd-sourced visual inspection

5526 real-world time series with labels for the location of all abrupt changes in level, variability, trend, presence/absence of data points, and irregular outliers. The time series were produced from a range of electronic health record data extracts from a large UK hospital group. Values in each data field were aggregated by day/week/month, and numeric summary values calculated for each timepoint from the (often non-numeric) data by applying simple functions (e.g. number of values present, percentage of missing values, number of distinct values, median value). Labels were produced by visual inspection of time series plots from ~2000 volunteers, via the Health Record Hiccups project on the Zooniverse platform (https://www.zooniverse.org/projects/phuongquan/health-record-hiccups). Volunteers drew a vertical line on the image wherever they saw a change point (green line if they were certain, yellow line if they were unsure). Consensus labels per image were calculated using density based clustering with noise (R v3.6.3, dbscan v1.1-5), and converted back to a date.本数据集包含5526条带标注的真实世界时间序列（time series），标注信息涵盖所有水平、变异性、趋势的突变位置，数据点的存在/缺失状态，以及不规则离群值的位置。该时间序列数据源自英国一家大型医院集团的多份电子健康记录（Electronic Health Record, EHR）提取结果。每个数据字段中的数值均按日、周、月粒度进行聚合，并通过简单函数从（通常为非数值型的）原始数据中计算得到每个时间点的数值汇总统计量，具体包括存在的数值数量、缺失值占比、不同值的个数、中位数等。标注信息由约2000名志愿者通过Zooniverse平台（https://www.zooniverse.org/projects/phuongquan/health-record-hiccups）上的“健康记录小插曲（Health Record Hiccups）”项目，对时间序列图表进行目视检视生成：志愿者在观测到突变点的位置绘制竖线，确认存在突变时使用绿色线条，存疑时使用黄色线条。后续研究团队使用基于R v3.6.3、dbscan v1.1-5实现的带噪声密度聚类算法，计算得到每张图表的共识标注，并将其转换为具体日期。