In the clinical setting the accepted method for measuring cardiac output (CO) has been for many years single bolus thermodilution using a pulmonary artery catheter (PAC). It has been available since the 1970s1 and has become the standard against which all new methods of CO monitoring are now judged.2–4
Newer, less-invasive methods of CO measurement are currently undergoing clinical evaluation, such as esophageal Doppler and arterial pulse wave contour. Validation studies for these newer techniques involve comparing readings from the new method with those from a reference method, usually thermodilution. Unfortunately, reports on the performance of these new technologies are conflicting and often flawed by poor experimental design and statistical analysis.2 Thus, several authors have recently called for a consensus on how such studies should be performed and analyzed.4,5 Furthermore, there have been calls for authors to clearly state the accuracy of their reference method.6
Knowing the accuracy, or precision error, of one's reference standard is of paramount importance when applying statistical methods and making judgments on the reliability of a new method of CO measurement, especially when there are sizable precision errors in the reference method of ±10%–20%, as with thermodilution.7 Currently, Bland and Altman limits of agreement of ±30% are used by most authors to show the acceptability of a new CO monitoring device,3 and these limits are based on a ±20% estimate for the precision error of the thermodilution reference method.7 However, ±20% may be an overestimate. More recent authors have suggested that a ±10% precision error may be more realistic.8,9 Hence, studies that will clarify the size of this precision error are needed.
We built a bench top test rig that circulated tap water through a loop of soft plastic aquarium tubing that had a diameter of approximately 1.6 cm. Flow rates of 0.5 to 10 L/min were generated by a constant rate electric aquarium water pump and variable orifice flow regulator. The flow remained constant and nonpulsatile at each regulator setting. The pump was placed in a 10-L container, which acted as a reservoir to prime the pump. The temperature within the tank and test rig was kept constant at 36.5°C by 2 heaters, thermostat, and water mixer (Fig. 1).
The flow rate through the test rig was measured by an externally placed ultrasonic transit flowprobe, size A16, and T106 small animal flowmeter (Transonic System, Ithaca, New York). Ultrasound gel was applied to the inside of the probe housing to provide an acoustic contact with the outside of the tubing. The probe was cleaned and gel was reapplied before each test session. However, the probe was not calibrated correctly for the thickness of the tubing wall, so the test rig needed to be calibrated before each test session by timed filling of a measuring cylinder.
Thermodilution measurements of the flow through the test rig were made by injecting boluses of 5 mL of ice-cold (0°C to 4°C) water (indicator) into the rig via the central venous pressure (CVP) port with the tip of the PAC placed downstream to the CVP lumen exit. The PAC was connected to a monitor that calculated CO from the area under the curve of temperature change detected by the thermistor at the PAC tip.
Two methods were used to place the PAC within the test rig. To replicate clinical measurement, we introduced the PAC into the lumen of the test rig via a small puncture hole in the wall of the tubing (Fig. 1). The distal end of the CVP lumen through which the indicator was injected was positioned within the test rig with the thermistor positioned downstream in the flow chamber. To prevent water leakage from the puncture hole, it was sealed with silicon glue.
However, a second alternative method of retrograde PAC placement was used when the performance of several PACs were compared during the same test session. In this method, cold indicator was injected into the test rig lumen via an independent port upstream of the flow chamber (Fig. 1), thus bypassing the CVP lumen and catheter dead-space. The PAC was passed retrograde into the flow chamber, against the direction of flow, through the distal end of the tubing, which emptied into the reservoir.
Reliability of the Test Rig
Reliability of the test rig was demonstrated by
- First, showing the agreement between measuring cylinder and flowmeter readings over a range of flows from 1 to 10 L/min. Readings taken at 0.5 L/min intervals were compared graphically. Typically, the regression line was Y = 1.54X – 0.24, where Y = flowmeter reading and X = timed cylinder reading, and it deviated from the line of identity (i.e., Y = X). Thus, there was a significant gain in the flowmeter readings of 1.54-fold. Therefore, the flowprobe and meter needed calibration before each use, and a correction factor derived from the slope of the regression line was used.
- Next, showing the precision of the flowmeter readings (after calibration) against cylinder readings by using a Bland and Altman analysis.10
- Finally, showing the repeatability of the flowmeter system readings over time by comparing the plots of flowmeter readings against cylinder collection times after 1, 2, 3, 8, and 12 hours.
The 2 main brands of PAC currently available in Hong Kong, Arrow (Arrow Int., Teleflex Medical, Reading, Pennsylvania) and Edwards Swan-Ganz (Edwards Lifesciences, Irvine, California) were studied using the test rig. Standard triple-lumen 7F catheters were used. These PACs did not measure continuous CO nor mixed venous oxygen saturation.
The performances of the PACs were first assessed using a Siemens SC9000 monitor with CO module (Siemens Medical Systems, Inc., Danvers, Massachusetts). Thermodilution measurements were made by injecting 5 mL of ice-cold saline using a standard PAC cooling coil and temperature measuring kit (Viggo-Spectramed Pte. Ltd., Singapore). Later the performance of 2 other CO monitors was tested, the Sirecust 1261 (Siemens Medical Systems Inc.) and the IntelliVue MP50 (Philips Medical Systems Inc., Andover, Massachusetts).
Data Collection and Analysis
The precision error of thermodilution CO was treated as having 2 distinct components: that due to random variations arising from the act of making a reading (i.e., interreading random error) and that due to systematic variations in the measurement equipment (i.e., between-catheters and monitoring system errors). The two sources of error were studied separately.
The Random Component
The random component of the precision error was investigated by taking repeated readings over a range of flow rates. A single catheter of each brand, Arrow and Edwards, was tested. To determine the influence of catheter dead-space on the size of this error, both the standard (with dead-space) and retrograde (without dead-space) PAC placement methods were evaluated. Ten readings were taken at a single flow rate. Further sets of 10 readings were then taken at 1 L/min intervals over a range of flow rates from 1 to 10 L/min (the uncorrected flowmeter reading). The Siemens SC9000 monitor was used to measure CO.
The random component of the precision error was determined by calculating its coefficient of variation (CV). First, the mean and SD for each set of 10 readings at a single flow rate was determined. The CV was calculated using the SD divided by the mean, which was expressed as a percentage (CV = (SD/mean)%). This was repeated for each flow rate, and then the average CV for all the CV measurements was determined.
To standardize our precision error measurement that was based on the CV that used 1 SD to the more commonly accepted presentation of 95% confidence intervals, we multiplied the average CV by ±1.96. Precision error in the manuscript will refer to 95% confidence intervals unless the CV value is specifically stated.
The Systematic Component
The systematic component of the precision error was investigated by plotting a regression, or calibration, line for a single PAC, measuring parameters that describe the line and repeating the process for several PACs. Such data were collected from 5 Arrow and 5 Edwards PACs. Only the retrograde method of PAC placement was used because the protocol required us to frequently change the PAC in the test rig. Five paired readings were collected at each of 5 flow rates between 0 to 10 L/min (the uncorrected flowmeter reading), and used to plot a calibration line for each PAC. By plotting the regression line for data over a range of flow rates, the random component of the precision error arising from the between-readings variation was eliminated, leaving only the systematic component.
From each plot of catheter performance a number of statistical variables that quantified its calibration line were determined:
- the slope or gradient (m);
- the correlation coefficient (r);
- the predicted value for a thermodilution reading (without any random component) when the flow in the rig was set at 5 L/min, which was calculated from each calibration line equation (Y = mX + c); and
- the deviation of the slope of the calibration line (m − 1) from the line of identity (Y = X).
The gradient of the calibration line can be used to quantify the systematic component of the precision error. Because the gradient is independent of flow rate, it provides a normalized estimate of systematic error that is unaffected by flow rate. The slope of each calibration line is formed by the hypotenuse of a right-angled triangle a, b, and c, for which a is the horizontal side and b is the vertical side. The slope or gradient (m) is simply b/a. When analyzing calibration lines, the important statistic is their deviation from the line of perfect agreement or identity Y = X, which has a gradient of unity (m = 1). The deviation (Δm) of the calibration line from this line (Δm) is (b/a) − 1. When deviations are small (i.e., <10°), a very simple relationship can be applied to the systematic error component and Δm in which the systematic error approximates to the value of Δm. Therefore, if Δm = 0.05, the systematic error is 5%, etc. The proof is based on the geometric relationship that Δm = tanΔθ = Δb/a for the small triangle and angle formed between the line of identity and the calibration line. To calculate the average divergence of the calibration lines, we used only absolute or positive values of Δm. The average divergence (mean(Δm)) for all catheters tested provided an estimate of the CV for the systematic error, which was expressed as a percentage. This was multiplied by 1.96 to provide the precision error component.
Difference in the Performance of 3 CO Monitors
Three different models of CO monitors were compared, Siemens SC9000, Sirecust 1261, and IntelliVue MP50. The experimental set-up and statistical analysis were the same as those described previously.
Reliability of the Test Rig
The correlation between flowmeter and cylinder measurements of test rig flow were excellent (r = 0.99; P < 0.0001), despite the large discrepancies in absolute values. The bias between pairs of readings after calibration was −0.0003 L/min and 95% confidence limits were −0.10 to +0.10 L/min. The repeatability of the flowmeter system readings over time was shown by comparing the plots of flowmeter readings against cylinder collection times after 1 hour and 2 hours. There was no discernable difference between plots. Similar results were obtained after 3, 8, and 12 hours.
The Random Variation Between Individual Thermodilution Measurements
The average CVs for the different flow rates were 5.4% and 4.8% for Arrow and Edwards PACs, respectively (Fig. 2). These average CVs were reduced to 2.7% and 3.8%, respectively, when the cold injectate did not pass through the catheter lumen and the dead-space was bypassed (P < 0.001; t test).
Comparison of Calibration Lines from Individual Catheters
The slope of the calibration lines from individual catheters varied (Fig. 3). The mean (range) predicted readings at a 5 L/min flow for Arrow PACs was 5.1 (4.9 to 5.2) L/min and for the Edwards PACs was 4.9 (4.6 to 5.2) L/min. The deviation (i.e., average CV [systematic]) of the slopes from the line of identity for Arrow PAC was ±5.8% and for Edwards PAC ±6.0%, and overall was ±5.9%. Although the calibration lines from the Arrow catheters were more uniform than were those from the Edwards catheters, they were not as close to the line of identity, lying slightly above it.
Precision Error Calculations
On the basis of our CV data, the component of precision error arising from the random errors for Arrow and Edwards PACs were ±10.6% and ±9.4%, respectively, and overall was ±10.0% (i.e., 95% confidence intervals). On the basis of our calibration line data, the systematic components of the precision error were ±11.4% and ±11.8%, respectively, and overall was ±11.6%. Therefore, the total estimated precision error for a single thermodilution reading, derived from CV(random) and CV(systematic) was ±15.3%, and this was reduced to ±13.0% for triplicate readings (i.e., CV = √[CVor CE(rand)2 + CV(syst)2]).
The Effect of Using Different CO Monitors
There were significant differences among the performances of the 3 CO monitors. The Siemens SC9000 monitor readings were more precise than were those of the other 2 monitors in respect to the interreading and between-catheters variability (Table 1). The precision errors for the Sirecust 1261 and the Philips MP50 monitors were 45% and 100% (or 1.45 and 2.0 times) more than those from the Siemens SC9000 monitor. The 5 Arrow catheters provided more consistent readings than did the 5 Edwards catheters. The calibration lines and their slopes were more consistent with the Siemens SC9000 monitor, but the Philips MP50 significantly underread (Table 1).
When using the Siemens SC9000 monitor to perform thermodilution measurements, our investigation found that the precision error (i.e., 95% confidence interval data) for PACs was ±15.3% for a single reading and ±13.0% for triplicate readings. The random component of this error was ±10.0% (or ±5.8% for triplicate) (CV: ±5.1% and ±2.9%) and the systematic component was ±11.6% (CV: ±5.9%). The Sirecust 1261 monitor increased the precision error by 45% and the Philips MP50 monitor by 100% (Table 1).
The reproducibility of serial thermodilution readings has been investigated previously in both in vitro models11–13 and clinical settings.14–18 Reproducibility in most of the studies was measured from serial thermodilution reading and quoted as the CV based on SD, rather than 95% confidence intervals, or precision error, as used today. For in vitro studies the CVs were about ±5% (range: ±2.5% to ±8.5%), which was in keeping with our Siemens 9000 result of ±5.1%. For in vivo or clinical testing the CVs were higher and ranged from ±4.8 to ±8.6%. However, only Fischer et al. have specifically measured both random and systematic errors arising from the catheter system.15 Using a statistical analysis based on repeatability data from paired thermodilution and dye dilution readings, they were algebraically able to derive the random and systematic error components using the adding-of-variances approach. These authors found a systematic component of the error for thermodilution of ±5% (CV), which was in keeping with our result of ±5.9%. For dye dilution the systematic error was ±15% (CV).
More up-to-date information on the accuracy of the current makes of thermodilution PAC is not readily available in the literature, and there have been calls to reappraise the current accepted precision error of thermodilution readings of ±20%.8,9 Work dating back before 1982 published by Stetz et al.7 is still used today to set the precision of the reference method thermodilution in most validation studies. Stetz et al. performed a statistical review of 9 papers published between 1968 and 1979 that provided CV reproducibility data on the performance of thermodilution and from these data determined the minimum acceptable percentage change in CO between a pair of single or triplicate sets of readings. Ambient temperature and volume of 10 mL were used for the injectate in most of these studies. From their meta-analysis, Stetz et al. recommended confidence limits for thermodilution readings of ±13% (triplicate readings) and ±22% (single reading), which are still in use today.7 Precision error data from our in vitro study of ±15.3% and ±13.0% would suggest that these limits are still close to the truth and therefore that the precision of the thermodilution method has not been improved over the last 20 to 30 years.
Our data also show that the choice of CO monitor can have a significant effect on the size of the precision error, which increased by >45% for the Sirecust monitor and by 100% for the Philips monitor. Thus, the precision error can vary greatly when using different monitoring systems, and this brings into question the choice of a standard precision error of ±20% for thermodilution measurement and the subsequent use of limits of agreement of less ±30% to test new CO device performance.3 In many published studies at present, the choice of monitor and catheter is not stated, or even worse, the reader does not know whether a variety of different monitors and catheters have been used.
The most likely origin for the systematic component of the precision error in our investigation is slight manufacturing differences between individual catheters. Pilot data comparing calibration lines when the tip of the PAC was placed in different positions within the flow chamber did not show such wide variations (unpublished validation data). Manufacturers do not provide this sort of quality control information about the performance of their PACs. Therefore our estimates of systematic errors provide an alternative source of such information. However, data from the Sirecust and Philips monitors (Table 1) showed an even wider variation in calibration lines, which suggests that there is an important issue regarding precision and the interface between different PAC and monitor systems. This needs further investigation.
In a recent paper, Cecconi et al. proved that without knowing the precision error of the reference method, it is difficult to interpret the meaning of the percentage error, or limits of agreement, from Bland and Altman analysis.6 Our results confirm that the precision error of the reference method can vary quite considerably between studies and thus stresses their message. However, their advice to measure the precision of the reference method at the bedside using serial readings and then calculate the reproducibility using the CV is not totally correct because only the random component, and not the systematic component, of the precision error is measured. Unfortunately, it is the only possible measurement of precision error that can easily be made at the bedside. Therefore, we suggest adding the best available estimate of the systematic component. On the basis of data from our investigation, this would be ±11.6%.
Our data come from an in vitro laboratory investigation, so caution is needed when applying them to a clinical setting. In vitro testing of thermodilution catheters may not mimic clinical use, but it does have a number of important advantages. Flow through the rig can be accurately determined to ±1%–2% (calibration data from our study). There is practically no limitation on the number of readings that can be taken, and thus a full range of flow rates can be tested, which in our investigation facilitated the plotting of calibration lines and estimating the systemic component. However, the model also had a number of important limitations. (i) Flow in the test rig was not pulsating and thus reading variability that would increase the random error component arising from the fluctuations in bloodflow was not present. (ii) Tap water rather than blood was used as the circulating fluid. The obvious problems of working with a biological fluid such as blood influenced our choice. However, blood differs from water by being a non-Newtonion fluid and having a lower heat capacity, 3.6 in comparison with 4.2 kJ/Kg/°K, respectively, a factor of 0.9. Differences in heat exchange between the cold injectate and circulating fluid can affect the size of the thermodilution curve and therefore the slope of the calibration line. Some authors add a correction factor into their calculation of CO to compensate.14 (iii) The tip of the PAC in vivo is normally placed in the pulmonary artery or one of its branches. In the test rig it was placed in a straight piece of 1.6-cm-diameter tubing, which may have also affected our results.
Another important cause of measurement error with thermodilution not addressed by our investigation was lung ventilation, which causes cyclical variations in cardiac output. Timing of the cold injection is therefore important when performing thermodilution. Whether readings should be performed at random times or synchronized with the ventilator cycle and its effect on precision would need consideration in any clinical trial.19
Catheter dead-space also plays an important role in determining PAC thermodilution precision error. A typical PAC is just over 1 m in length and has 3 lumens. One of these lumens, the CVP port, is used to inject the cold indicator and has a volume of approximately 1 mL, which is referred to as dead-space. This dead-space contributes significantly (i.e., ±10%–20%) to the volume of the injectate (i.e., 5 to 10 mL). Indeed if the temperature of the fluid in the dead-space is different from that of the indicator, then the thermal effect and thus the measurement of CO is altered. Because the temperature of the dead-space fluid is difficult to control, it becomes a major source of variability in CO readings. Heat gain by the injectate as it passes through the dead-space may also be a factor and can be influenced by (a) injection time, which should be less than 4 seconds, (b) length of catheter placed IV, (c) the temperature of the blood, and (d) the hemoglobin level.12 Minimizing the effect of dead-space therefore becomes an important consideration in PAC manufacture and use. The impact of dead-space on precision was shown in the first part of our investigation, in which we compared the CV of measurements with and without it. The presence of dead-space added up to ± 5% to the total precision error for a single reading (Table 1).
Three different design layouts were described in the literature for building a test rig for thermodilution catheter testing. Designs involved either constant or pulsatile flow systems. Constant flow systems were based on either a constant water pressure with flow regulator20 or a variable rate roller pump.11,21 Some authors developed a piston pump system with unidirectional valves to mimic the pumping action of the heart, which they used to investigate left ventricular ejection fraction catheters with a high- response-rate thermistor.22,23 Most designs were based on circulating water. Only 1 used human blood11 and 1 used glycine.12 Flow rate was most commonly measured by timed measuring cylinder filling. The temperature of the water, or blood, was kept constant at 36°C to 38°C by a thermostated heater and water mixer. In our design, we used a constant flow of water, rather than pulsatile, for ease of construction. A heavy-duty aquarium pump was used in preference to a roller pump.
In the present investigation we chose an injectate volume of 5 mL of ice-cold water at 0°C to 4°C. Ice-cold water provides a greater thermal challenge (2°C to 37°C) than does room- temperature water (22°C to 37°C), and thus the amplitude of the thermodilution curve is larger and the readings more accurate. However, injectate volumes above 5 mL are reported to have little influence on CV and precision of readings.19 Larger volumes may even reduce accuracy because of the recirculation of the indicator.24 In our test rig, injectate volumes of 5 and 10 mL were shown to have a similar CV (±3.1% and ±2.7%) (P = 0.37), whereas a 3-mL injectate had a larger CV (±7.7%) (P = 0.02) (unpublished validation data).
Significant random and systematic errors occur during thermodilution measurement, and both should be included in any estimate of the precision error. Although the random component can be simply estimated from serial readings, the systematic component is much more difficult to measure. According to Cecconi et al., an estimate of precision error should be determined before performing every validation study to set evidence-based acceptance criteria.6 However, the choice of catheter and monitoring set-up can also have a significant impact on the size of this precision error. On the basis of data from our in vitro investigation, our best estimate (95% confidence interval data) for the random (interreading) error is 10% for single readings and ±5.8% for triplicate readings, and the systematic (between catheters) error is ±11.6%. Thus, the overall precision error is ±15.3% for a single reading, and ±13.0% for triplicate readings. However, different equipment can increase this precision error by 1.5 to 2.0 times or over 50%–100%. Furthermore, one should be cautious before laboratory-acquired data are applied to the clinical setting, because these may also be additional clinical and equipment sources of measurement error. Reevaluation of the current acceptance criteria of ±30% is needed in light of our new data, but first we need to confirm our findings by repeating the investigation with a pulsating model and animal testing.
We thank Mr. Kafai Mak for his drawing of the test rig (Fig. 1).
1. Swann HJC, Ganz W, Forrester J, Marcus H, Diamond G, Chonette D. Catheterization of the heart in man with use of a flow-directed balloon-tipped catheter. N Engl J Med 1970;283:447–51
2. Cecconi M, Grounds M, Rhodes A. Methodologies for assessing agreement between two methods of clinical measurement: are we as good as we think we are? Curr Opin Crit Care 2007;13:294–6
3. Critchley LA, Critchley JA. A meta-analysis of studies using bias and precision statistics to compare cardiac output measurement techniques. J Clin Monit Comput 1999;15:85–91
4. Linton NW, Linton RA. Is comparison of changes in cardiac output, assessed by different methods, better than only comparing cardiac output to the reference method? Br J Anaesth 2002;89:336–7
5. Bein B, Renner J, Scholz J, Tonner PH. Comparing different methods of cardiac output determination: a call for consensus. Eur J Anaesthesiol 2006;23:710
6. Cecconi M, Rhodes A, Poloniecki J, Della Rocca G, Grounds RM. Bench-to-bedside review: the importance of the precision of the reference technique in method comparison studies—with specific reference to the measurement of cardiac output. Crit Care 2009;13:201
7. Stetz CW, Miller RG, Kelly GE, Raffin TA. Reliability of the thermodilution method in the determination of cardiac output in clinical practice. Am Rev Respir Dis 1982;126:1001–4
8. Berthelsen PG, Eldrup N, Nilsson LB, Rasmussen JP. Thermodilution cardiac output cold vs room temperature injectate and the importance of measuring the injectate temperature in the right atrium. Acta Anaesthesiol Scand 2002;46:1103–10
9. Nilsson LB, Nilsson JC, Skovgaard LT, Berthelsen PG. Thermodilution cardiac output—are three injections enough? Acta Anaesthesiol Scand 2004;48:1322–7
10. Bland JM, Altman DG. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 1986;1:307–10
11. Bilfinger TV, Lin CY, Anagnostopoulos CE. In vitro determination of accuracy of cardiac output measurements by thermal dilution. J Surg Res 1982;33:409–14
12. Forrester JS, Ganz W, Diamond G, McHugh T, Chonette DW, Swan HJ. Thermodilution cardiac output determination with a single flow-directed catheter. Am Heart J 1972;83:306–11
13. Hoel BL. Some aspects of the clinical use of thermodilution in measuring cardiac output. With particular reference to the Swan–Ganz thermodilution catheters. Scand J Clin Lab Invest 1978;38:383–8
14. Andreen M. Computerized measurement of cardiac output by thermodilution: methodological aspects. Acta Anaesthesiol Scand 1974;18:297–305
15. Fischer AP, Benis AM, Jurado RA, Seely E, Teirstein P, Litwak RS. Analysis of errors in measurement of cardiac output by simultaneous dye and thermal dilution in cardiothoracic surgical patients. Cardiovasc Res 1978;12:190–9
16. Kohanna FH, Cunningham JN Jr. Monitoring of cardiac output by thermodilution after open-heart surgery. J Thorac Cardiovasc Surg 1977;73:451–7
17. Olsson B, Pool J, Vandermoten P, Varnauskas E, Wassén R. Validity and reproducibility of determination of cardiac output by thermodilution in man. Cardiology 1970;55:136–48
18. Vandermoten P, Bernard R, de Hemptinne J, Gillet JM, Lenaers A. Cardiac output monitoring during the acute phase of myocardial infarction: accuracy and precision of the thermodilution method. Cardiology 1977;62:291–5
19. Reuter DA, Huang C, Edrich T, Shernan SK, Eltzschig HK. Cardiac output monitoring using indicator dilution techniques: basics, limits, and perspectives. Anesth Analg 2010;110:799–811
20. Rubini A, Del Monte D, Catena V, Attar I, Cesaro M, Soranzo D, Rattazzi G, Alati GL. Cardiac output measurement by the thermodilution method: an in vitro test of accuracy of three commercially available automatic cardiac output computers. Intensive Care Med 1995;21:154–8
21. Norris SL, King EG, Grace M, Weir B. Thermodilution cardiac output—an in vitro model of low flow states. Crit Care Med 1986;14:57–9
22. Maruschak GF, Schauble JF. Limitations of thermodilution ejection fraction: degradation of frequency response by catheter mounting of fast-response thermistors. Crit Care Med 1985;13:679–82
23. Santos I, Ferreira da Rocha A, Assis de O Nascimento F, Souza Neto J, Valvano JW. Measurement of ejection fraction with standard thermodilution catheters. Med Engl Phys 2002;24:325–35
24. Runciman WB, Ilsley AH, Roberts JG. Thermodilution cardiac output—a systematic error. Anaesth Intensive Care 1981;9: 135–9