1、Estimating Cardiac Output from Arterial Blood Pressure Waveforms:a Critical Evaluation using the MIMIC II DatabaseJX Sun, AT Reisner, M Saeed, RG MarkHarvard-MIT Division of Health Sciences and Technology, MIT, Cambridge, MA, USAAbstractCardiac output (CO) estimation using arterial bloodpressure (AB
2、P) waveforms has been an active area ofphysiology research over the past century. However, the ef-fectiveness of the estimators has not been extensively stud-ied in a clinical setting. In this paper, we evaluate 11 well-known COestimatorsusingclinical radial ABP waveformsfrom the Multi-Parameter Int
3、elligent Monitoring for Inten-sive Care II (MIMIC II) database, using thermodilutionCO (TCO) as reference for comparison. We compare esti-mations to 988 TCO measurements in 84 patients, totaling165 hours of ABP waveforms sampled at 125 Hz. As a nec-essary step for producing absolute CO estimates, we
4、 alsopresent 3 methods of calibrating the estimators, each tai-lored towards a different use model. The results show thatthe standard deviation of error between TCO and the bestCO estimators is approximately 1 L/min for absolute COestimates. For relative estimates without calibration, thebest CO est
5、imator has 18% error at 1 standard deviation.1. IntroductionCardiac output (CO) is a key parameter in assessing cir-culatory function. The nominal CO value for a healthyhuman is about 5 liters per minute. Currently in clinicalpractice, the gold standard for CO measurement is ther-modilution CO (TCO)
6、, which involves the insertion of aSwan-Ganz catheter into the pulmonary artery. Admin-istered primarily in intensive care units (ICUs), TCO isusually measured intermittently, is very invasive, and maycause severe complications. It would be a tremendous as-set to healthcare if one could determine CO
7、 accurately, re-liably, and continuously using less invasive, indirect meth-ods. Indeed, in the past century, over a dozen schemeshave been proposed and developed to estimate CO usingarterial blood pressure (ABP) waveforms obtained from apatientfarlessinvasively. Someoftheseestimatorsrelyonelaborate
8、 models of the heart and vasculature while othersuse artificial intelligence methods such as pattern match-ing and classification trees. The published estimators havenot been extensively evaluated with a large set of clinicalABP waveforms, hence the performance of CO estimationis still uncertain. St
9、udies in the past have mostly been con-ducted on a smallset of subjects under well-controlled lab-oratory conditions. It is entirely possible that there will becircumstancesinrealworldclinicalpracticeinwhichtheseindirect methods produce inaccurate estimates.The Multi-Parameter Intelligent Monitoring
10、 for Inten-sive Care II (MIMIC II) database has physiologic wave-form data from over 3500 ICU patients hospitalized atBeth Israel Deaconess Medical Center, Boston, USA. Thedatabase has about 100 patient records that contain ABPwaveforms and TCO measurements simultaneously. Ourgoal is to evaluate 11
11、of the CO estimators on a suitablesubset of these patients using TCO as reference standard.Table 1. CO estimatorsEstimator CO = k belowMean Pressure PmeanWindkessel 1 Ppulse HRSystolic Area 2 Asys HRWarner Time Correction 3AG1 + TsysTdiasAHAsys HRLiljestrand thus, we cannot calibrate to produce an a
12、bsolute CO esti-mate. However, it is still useful to know relative fractionalchanges in CO. For example, if an uncalibrated estima-tor output decreased from 4000 to 2000, we would like toknow if the true CO has decreased by a similar fractionalamount. For evaluation, we define percentage changes inT
13、CO and each uncalibrated CO estimator as:X = xxx R = rrrr and x are averages in TCO and CO estimator output,respectively. To report error, we examine the differencebetween X and R. For example, if X = 0.3 and R = 0.4,the magnitude of error would be reported as 0.1, or 10%.3.4. Patient selectionWe wa
14、nt to evaluate the CO estimators on patients thathave relatively clean ABP waveforms and a significantnumber of TCO measurements. We accept patients in theMIMIC II database if all of the following are true:1. SQI flags 20% of beats in ABP waveform.2. Patients with 5 TCO measurements.3. Patients that
15、 donot have intra-aorticballoon pumpsor with abnormal aortic or tricuspid valve function.Based on these criteria, 84 patients were identified, eachaveraging 12 TCO measurements. Some statistics for thepopulation is listed in Table 2.Table 2. Subject population statisticsParameter Mean RangeAge years
16、 70 4095TCO L/min 5.3 212TCO per patient L/min 2.5 1.564. Results and discussionThe first 3 columns of Table 3 list the standard devia-tion of error in liters per minute between TCO and eachestimator for the 3 different calibrations for absolute COestimates. The last column of the table lists the pe
17、rcentageerror at 1 standard deviation for relative CO estimates. The95% confidence intervals are about twice the values in thetable. Figure 5 shows a Bland-Altman plot for the Liljes-trand 12:147378.2 WesselingK,NicolsW,WitB,H.W. Abeat-to-beatcardiacoutput computer. In Proceedings of the third Inter
18、nationalConference on Medical Physics. 1972; .3 Warner H, Swan H, Connolly D, Tompkins R, Wood E.Quantitation of beat-to-beat changes in stroke volume fromthe aortic pulse contour in man. Journal of Applied Physi-ology 1953;5:495.4 Liljestrand G, Zander E. Experimental Medicine 1940;59:105.5 Herd J,
19、 Leclair N, Simon W. Arterial pressure pulse con-tours during hemorrhage in anesthetized dogs. Journal ofApplied Physiology 1966;21.6 Wesseling K, Wit B, Weber J, Smith N. A simple device forthe continuous measurement of cardiac output. AdvancedCardiovascular Physiology 1983;5.7 Godje O, Lamm P, Sch
20、mitz C, Theil M, B. R. Continuous,lessinvasive, hemodynamicmonitoringinintensivecareaf-ter cardiac surgery. Thoracic Cardiovascular Surgery 1998;46:242249.8 Jonas M, Hett D, Morgan J. Real time, continuous moni-toring of cardiac output and oxygen delivery. InternationalJournal of Intensive Care 2002
21、;9.9 Bourgeois J, Gilbert B, Donald D, Wood E. Characteristicsof aortic diastolic pressure decay with application to con-tinuous monitoring of changes in peripheral vascular resis-tance. Circulation Research 1974;35:5666.10 Wesseling K, Jansen R, Settels J, Schreuder J. Computa-tions of aortic flow
22、from pressure in humans using a non-linear, three-element model. Journal of Applied Physiology1993;74.11 Zong W, Heldt T, Moody G, Mark R. An open-source al-gorithm to detect onset of arterial blood pressure pulses. InComputers in Cardiology 2003. Los Alamitos: IEEE Com-puter Society Press, 2003; 259262.12 Stetz C, Miller R, Kelly G, Raffin T. Reliability of the ther-modilution method in the determination of cardiac outputin clinical practice. Am Rev Respir Dis 1982;126.Address for correspondence:James X. SunMIT, Room E25-505, 77 Mass Ave, Cambridge, MA 02139,USAxinsunmit.edu
