Liu Nehemiah T, Kramer George C, Khan Muzna N, Kinsky Michael P, Salinas José
From the US Army Institute of Surgical Research (N.T.L., J.S.), Fort Sam Houston, San Antonio; and Resuscitation Research Laboratory (G.C.K., M.N.K., M.P.K.), Department of Anesthesiology, University of Texas Medical Branch, Galveston, Texas.
J Trauma Acute Care Surg. 2015 Oct;79(4 Suppl 2):S85-92. doi: 10.1097/TA.0000000000000671.
This study was a first step to facilitate the development of automated decision support systems using cardiac output (CO) for combat casualty care. Such systems remain a practical challenge in battlefield and prehospital settings. In these environments, reliable CO estimation using blood pressure (BP) and heart rate (HR) may provide additional capabilities for diagnosis and treatment of trauma patients. The aim of this study was to demonstrate that continuous BP and HR from the arterial BP waveform coupled with machine learning (ML) can reliably estimate CO in a conscious sheep model of multiple hemorrhages and resuscitation.
Hemodynamic parameters (BPs, HR) were derived from 100-Hz arterial BP waveforms of 10 sheep records, 3 hours to 4 hours long. Two models (mean arterial pressure, Windkessel) were then applied and merged to estimate COVS. ML was used to develop a rule for identifying when models required calibration. All records contained 100-Hz recording of pulmonary arterial blood flow using Doppler transit time (COFP). COFP and COVS were analyzed using equivalence tests and Bland-Altman analysis, as well as waveform and concordance plots.
Baseline COFP varied from 3.0 L/min to 5.4 L/min, while posthemorrhage COFP varied from 1.0 L/min to 1.8 L/min. A total of 315,196 pairs of data were obtained. Equivalence tests for individual records showed that COVS was statistically equivalent to COFP (p < 0.05). Smaller equivalence thresholds (<0.3 L/min) indicated an overall high COFP accuracy. The agreement between COFP and COVS was -0.13 (0.69) L/min (Bland-Altman). In an exclusion zone of 12%, trending analysis found a 92% concordance between 5-minute changes in COFP and COVS.
This study showed that CO can be reliably estimated using BPs and HR from the arterial BP waveform in combination with ML. A next step will be to test this approach using noninvasive BPs and HR.
本研究是推动利用心输出量（CO）开发用于战伤救治的自动化决策支持系统的第一步。此类系统在战场和院前环境中仍然是一项实际挑战。在这些环境中，利用血压（BP）和心率（HR）进行可靠的心输出量估计可为创伤患者的诊断和治疗提供额外能力。本研究的目的是证明，来自动脉血压波形的连续血压和心率结合机器学习（ML）能够在多出血和复苏的清醒绵羊模型中可靠地估计心输出量。
从10份绵羊记录的100赫兹动脉血压波形中获取血流动力学参数（血压、心率），记录时长为3至4小时。然后应用并合并两种模型（平均动脉压、风箱模型）来估计心输出量（COVS）。利用机器学习制定一条规则，以识别模型何时需要校准。所有记录均包含使用多普勒渡越时间对肺动脉血流进行的100赫兹记录（COFP）。使用等效性检验、布兰德-奥特曼分析以及波形和一致性图对COFP和COVS进行分析。
基线COFP在3.0升/分钟至5.4升/分钟之间变化，出血后COFP在1.0升/分钟至1.8升/分钟之间变化。共获得315,196对数据。对个体记录的等效性检验表明，COVS在统计学上等同于COFP（p < 0.05）。较小的等效性阈值（<0.3升/分钟）表明总体上COFP准确性较高。COFP与COVS之间的一致性为-0.13（0.69）升/分钟（布兰德-奥特曼分析）。在12%的排除区内，趋势分析发现COFP与COVS的5分钟变化之间的一致性为92%。
本研究表明，结合机器学习，利用动脉血压波形中的血压和心率能够可靠地估计心输出量。下一步将使用无创血压和心率来测试这种方法。
