Sub-Cohorts of MIMIC-III Waveform Database Matched Subset for Generalizability Benchmarking under Varying Distribution Shift in Clinical Time-Series Forecasting

The Medical Information Mart for Intensive Care (MIMIC) waveform databases are a popular, publicly available family of databases that provide patient monitoring recordings from ICUs. The MIMIC III waveform database matched subset v1.0 contains high-resolution bedside monitoring recordings of more than 10,000 patients recorded at the Beth Israel Deaconess Medical Centre in Boston, Massachusetts, between 2001 and 2012. The records of the MIMIC III waveform database matched subset v1.0 can be matched to patients and stays in the MIMIC III Clinical Database v1.4, which contains further information about a patient’s stay, such as performed procedures and administered medications. We chose the MIMIC III waveform database matched subset v1.0 due to its size, public availability, provision of recordings at 1-minute resolution or higher, and the extensive context of the clinical database. From the MIMIC III waveform database matched subset v1.0, a base dataset is constructed by including only records with at least 65 minutes of CVP, HR, SpO2, and systolic, diastolic, and mean ABP recordings. The recordings of the base dataset were pre-processed, inspired by Hinrichs et. al.. First, mean downsampling to 1/minute resolution ensures uniform recording resolution. Then data cleaning is applied, by removing any vital parameters that lie outside of physiologically plausible ranges (CVP: -15 to 50 mmHg; HR: 10 to 330 beats per minute; SpO2: 40 to 100%; systolic ABP: 8 to 350 mmHg; diastolic ABP: 8 to 150 mmHg; mean ABP: 8 to 180 mmHg), and all sequences where at least one vital parameter misses 15 minutes of recordings are removed. At this point, the recordings are windowed using a sliding window approach, with an observation window of 60 minutes, a prediction window of 5 minutes, and a step size of 5 minutes. The resulting samples are split into 80\% for the training set and 20\% for the test set, ensuring no patient leakage between the sets. Next, missing data is imputed by an iterative imputer. Finally, mean downsampling to a 1-per-5-minute resolution is performed. The diversity and size of the MIMIC dataset allow for a clinically meaningful definition of subsets that represent different degrees of distributional shifts between pairs. Two dataset pairs are based on whether a patient is admitted to the ICU ward after a specific surgery; the third pair is based on whether the data were recorded during mechanical ventilation; and the final dataset pair is based on administration of a class of medications used for blood pressure regulation, called vasopressors. The final scenarios are defined as follows: (C) Cardiac Surgery: All recordings with a procedure with an ICD-9 code between 35 and 39 or a CPT code between 33016 and 37799 during the hospital stay. It is assumed that if only one ICU stay is recorded for a hospital stay with a surgical procedure, then that ICU stay is post-surgical; therefore, all matching recordings with more than one ICU stay are excluded. (NC) No Cardiac Surgery: All recordings that do not have a procedure of ICD 9 code 35 to 39 or CPT code 33016 to 37799 during the associated hospital stay. (R) Respiratory Surgery: All recordings with at most one ICU stay, where during the hospital stay a procedure with an ICD 9 code between 30 to 34 or a CPT code between 30000 to 33000 or between 39000 to 39600 was performed. (NR) No Respiratory Surgery: All recordings that do not have a procedure with an ICD9 code between 30 and 34 or a CPT code between 30000 and 33000 or between 39000 and 39600 during the associated hospital stay. (M) Mechanical Ventilation: All days of a recording for which a procedure with CPT code 94003 is registered. However, the timespans of the last day of a sequence of days with an associated CPT code 94003 are excluded, since the MIMIC III base database records mechanical ventilation only on a day basis; the exact time point when ventilation was stopped is not recorded. For the same reason, days with CPT code 94002 are excluded, as this code indicates the start of mechanical ventilation. (NM) No Mechanical Ventilation: All timespans of a recording, for which no CPT code associated with ventilation (94002 or 94003) is associated. (V) Vasopressors: All timespans of records for which continuous administration of at least one vasopressor is recorded. Vasopressors are defined as: dobutamine, dopamine, epinephrine(-k), levophed(-k), vasopressin and neosynephrine(-k). (NV) No Vasopressors: All records for which no one-time or continuous medication administration of any vasopressor is recorded.重症监护医疗信息集市（Medical Information Mart for Intensive Care，下文简称MIMIC）波形数据库是一类广受欢迎且可公开获取的数据库集群，收录了重症监护室（intensive care unit，下文简称ICU）的患者监护记录。MIMIC III波形数据库匹配子集v1.0收录了2001年至2012年间，位于马萨诸塞州波士顿的贝斯以色列女执事医疗中心（Beth Israel Deaconess Medical Centre）采集的逾1万名患者的高分辨率床边监护记录。该子集的记录可与MIMIC III临床数据库v1.4中的患者及住院信息进行匹配，该临床数据库包含患者住院期间的更多相关信息，例如实施的诊疗操作与给药记录。 我们选用MIMIC III波形数据库匹配子集v1.0，原因在于其样本规模可观、可公开获取、支持1分钟及更高分辨率的监护记录采集，且配套临床数据库提供了丰富的上下文信息。我们从MIMIC III波形数据库匹配子集v1.0中构建基础数据集，仅保留同时包含至少65分钟中心静脉压（Central Venous Pressure, CVP）、心率（Heart Rate, HR）、脉搏血氧饱和度（Pulse Oximetry, SpO2）、收缩压、舒张压及平均动脉压（Arterial Blood Pressure, ABP）记录的样本。 本基础数据集的监护记录参考Hinrichs等人的研究进行预处理。首先通过均值下采样将分辨率统一至1次/分钟，确保所有记录的采样分辨率一致。随后执行数据清洗操作：移除超出生理合理范围的生命体征参数（中心静脉压：-15~50 mmHg；心率：10~330次/分钟；脉搏血氧饱和度：40%~100%；收缩压：8~350 mmHg；舒张压：8~150 mmHg；平均动脉压：8~180 mmHg），同时剔除任意一项生命体征存在15分钟及以上数据缺失的监护序列。 完成上述步骤后，采用滑动窗口法对监护记录进行分窗处理：观测窗口长度为60分钟，预测窗口长度为5分钟，步长为5分钟。将生成的样本按80%、20%的比例划分为训练集与测试集，确保两组样本间不存在患者数据泄露问题。接下来，采用迭代插补法对缺失数据进行填充。最后，通过均值下采样将分辨率进一步统一至1次/5分钟。 MIMIC数据集的多样性与规模支持我们基于临床意义定义多组子集，用于表征不同程度的跨子集分布偏移。其中两组子集对基于患者是否在特定手术后收入ICU病房；第三组子集对基于监护记录是否采集于机械通气期间；最后一组子集对基于是否使用了一类用于血压调节的药物——血管升压药（vasopressors）。 最终的划分场景定义如下： (C) 心脏外科手术：所有住院期间实施了国际疾病分类第9版（International Classification of Diseases 9th Revision, ICD-9）编码介于35~39之间，或当前操作术语编码（Current Procedural Terminology, CPT）介于33016~37799之间的诊疗操作的监护记录。假设若某手术住院期间仅记录了一次ICU住院，则该次ICU住院属于术后住院；因此，剔除存在多次ICU住院记录的匹配样本。 (NC) 非心脏外科手术：所有住院期间未实施ICD-9编码35~39或CPT编码33016~37799对应的诊疗操作的监护记录。 (R) 呼吸外科手术：所有住院期间实施了ICD-9编码介于30~34之间，或CPT编码介于30000~33000、39000~39600之间的诊疗操作，且仅存在一次ICU住院记录的监护记录。 (NR) 非呼吸外科手术：所有住院期间未实施ICD-9编码30~34或CPT编码30000~33000、39000~39600对应的诊疗操作的监护记录。 (M) 机械通气：所有记录了CPT编码94003对应诊疗操作的监护日期。但由于MIMIC III基础数据库仅按天记录机械通气信息，无法获取通气停止的精确时间点，因此剔除连续使用CPT编码94003的日期序列中的最后一天；同理，CPT编码94002代表机械通气启动，相关日期也一并剔除。 (NM) 非机械通气：所有未关联任何通气相关CPT编码（94002或94003）的监护时段。 (V) 血管升压药给药：所有记录了至少一种血管升压药持续给药的监护时段。血管升压药定义为：多巴酚丁胺（dobutamine）、多巴胺（dopamine）、肾上腺素（epinephrine(-k)）、去甲肾上腺素（levophed(-k)）、加压素（vasopressin）及去氧肾上腺素（neosynephrine(-k)）。 (NV) 未使用血管升压药：所有未记录任何一次性或持续性血管升压药给药的监护记录。