Accompanying data for "A Multilevel statistical process control chart framework for personalized patient monitoring: A simulation study using patients historical data"

Statistical Process Control (SPC) charts offer a potentially effective means of continuously monitoring patient time-series data to enable the early detection ofclinical abnormalities. A significant challenge in implementing SPC methods, such as the Exponentially Weighted Moving Average (EWMA), is obtaining sufficient phase-I data from individual patients. This data is crucial for accurately estimating control limits using the expected mean and standard deviation duringstable periods. In contrast to other fields where ample historical data is available, patient monitoring systems often lack enough observations to effectively set these limits. To address this issue, we propose a multilevel SPC approach that uses historical data from stable subjects. By fitting a multilevel model with arandom intercept, our method enables comparison of an individual’s longitudinal measurements against cross-sectional data from other stable patients. Themodel recursively estimates the expected mean within a calibration window, applying distinct control limit coefficients compared to phase-II monitoring period. Person-specific deviations, computed as residuals between actual and estimated values are then monitored via an EWMA control chart. Simulation studies usingsynthetic data modeled on left ventricular assist device (LVAD) patients were conducted to evaluate this approach. The model was evaluated across multiplesimulation scenarios, varying parameters such as sampling frequency, shift size and type, to assess its robustness under different operating conditions. Resultsshowed that longer calibration periods enhanced alignment with patient data, thereby increasing the true-positive rate and reducing false alarms. However, if the calibration window becomes excessively long, it may lead to overfitting, which could delay detection or result in an increased false-positive rate. Our multilevel SPC framework demonstrates potential to improve early detection of mean shifts in patient monitoring data, offering a refined approach to personalized healthcaremonitoring and data-driven clinical decision-making. This repository contains the simulated LVAD data used in our study on multilevel statistical process control for personalized patient monitoring. The simulations were generated using the CPAR (probability autoregressive) model across all parameter combinations. Data with shifts stored in data_with_shifts.zip file. Using the CPAR framework, we generated 25,600 distinct simulation scenarios (5 × 4 × 4 × 2 × 2 × 2 × 2 × 5 × 4 combinations of settings basedon the simulation nstudy). For each scenario, a cohort of 500 unique patients was created: 250 patients for monitoringand 250 patients from a stable cohort, randomly sampled to assess the impact of the simulation parameters. Each patient trajectory was generated uniquely for that specific scenario based on the CPAR model. Code can be accessed via HERE.统计过程控制（Statistical Process Control，SPC）是一种具备潜在应用价值的有效手段，可用于持续监测患者时序数据，实现临床异常状况的早期识别。在应用指数加权移动平均（Exponentially Weighted Moving Average，EWMA）这类SPC方法时，面临的一项显著挑战是难以获取足量的单患者阶段I数据——这类数据对于基于稳定时期的期望均值与标准差精准估算控制限至关重要。与其他拥有充足历史数据的领域不同，患者监测系统往往缺乏足够的观测样本以有效设定这些控制限。 为解决这一问题，本文提出一种多级SPC方法，该方法可利用稳定受试者的历史数据开展建模。通过拟合带有随机截距的多级模型，我们的方法能够将个体的纵向测量结果与其他稳定患者的横截面数据进行对比。该模型可在校准窗口内递归估算期望均值，并采用与阶段II监测周期不同的专属控制限系数。随后，以实际观测值与估算值的残差计算得到的个体特异性偏差，将通过EWMA控制图进行监测。 本研究采用基于左心室辅助装置（left ventricular assist device，LVAD）患者数据构建的合成数据开展仿真实验，以评估所提方法的性能。我们在多种仿真场景下对模型进行测试，通过调整采样频率、偏移幅度与偏移类型等参数，评估模型在不同工况下的鲁棒性。实验结果表明，更长的校准周期能够提升与患者数据的匹配度，进而提高真阳性率并减少虚假警报。但若校准窗口过长，则可能引发过拟合问题，导致异常检测延迟或假阳性率升高。我们提出的多级SPC框架展现出改善患者监测数据中均值偏移早期检测的潜力，可为个性化医疗监测与数据驱动的临床决策提供更为精准的解决方案。 本仓库包含了我们针对个性化患者监测的多级统计过程控制研究中所使用的仿真LVAD数据集。所有参数组合下的仿真数据均通过概率自回归（probability autoregressive，CPAR）模型生成，带偏移的数据存储于data_with_shifts.zip压缩文件中。 借助CPAR框架，我们共生成了25600种独立的仿真场景（基于本次仿真研究的参数设置，共包含5×4×4×2×2×2×2×5×4种组合）。针对每种场景，我们构建了包含500名独特患者的队列：其中250名用于监测，另外250名来自稳定队列，通过随机抽样以评估仿真参数的影响。每名患者的监测轨迹均基于对应场景下的CPAR模型单独生成。代码可通过HERE获取。