Introduction

Changes to ventilator settings are part of the routinemanagement of mechanically ventilated patients on theintensive care unit (ICU). ICU physicians make thesechanges, based on their pathophysiological knowledgeand experience, in an attempt to attain optimal arterialgas tensions and pH while avoiding the risks ofbarotrauma [1, 2], oxygen toxicity [3, 4], hypoxaemia,acidaemia and alkalaemia. Physiological simulationshave provided predictions of blood gas responses to in-terventions in mechanical ventilation [5, 6], althoughno comparison of their efficacy with that of physicianshas been published. We have developed a computer-based physiological simulator: The Nottingham Physiol-ogy Simulator (NPS). To have value as a clinical predic-tive tool, a simulator needs to possess greater accuracy(smaller bias) and better consistency (smaller 95% lim-its of agreement) than physicians. The aim of this studywas to compare the ability of the NPS and ICU consul-

tant (specialist) physicians to predict changes in arterialoxygen tension (PaO2), arterial carbon dioxide tension(PaCO2) and arterial pH after alterations were made topatients' ventilator settings.

Materials and methods

The NPS is a computer simulation of original multi-compartmentalrespiratory, acid-base, cardiovascular and cerebrovascular physio-logical models. Physiological variables may be entered into thesimulator in order to `align' it to a real patient. Details of this pro-cess are described elsewhere [6±8] and are described briefly in theappendix. The NPS has been shown to predict accurately changesin blood gas parameters following changes in ventilator settings[6].

The local ethics committee approved the study and agreed thatwritten consent was not necessary, as the study was observational.We collected data sets from mechanically ventilated patients be-fore and 10 min after physicians made changes in ventilator set-tings as part of the patients' normal management. All patientswere sedated and had no spontaneous ventilatory minute volume.

N.M. BedforthJ.G. Hardman

Predicting patients' responsesto changes in mechanical ventilation:a comparison between physiciansand a physiological simulator

Received: 29 September 1998Final revision received: 23 February 1999Accepted: 23 April 1999

N.M.Bedforth × J. G. Hardman ())University Department of Anaesthesia &Intensive Care, University Hospital,Nottingham, NG7 2UH, UKe-mail: jon.hardman@nottingham.ac.ukTel. + 44(1 15)9709229;Fax + 44(115)970 0739

Abstract We compared the accu-racy and reliability of a validated,physiological simulator and six in-tensive care specialists in predictingchanges in arterial oxygen tension(PaO2), arterial carbon dioxide ten-sion (PaCO2) and pH following ad-justment of mechanical ventilation.Twenty-five data sets were collectedbefore and after routine alterationsin ventilator settings. Fractional in-spired oxygen was adjusted in allpatients and minute volume was ad-justed in 13 patients. The simulatorwas more accurate and consistentthan all the physicians in predictingthe magnitude of PaO2 and pH

change. The simulator had a largerbias in estimating the magnitude ofchange of PaCO2 than four of thephysicians, but was more consistentthan all but one of the physicians.The simulator may prove to be auseful tool in the management ofmechanical ventilation. Incorpora-tion into mechanical ventilators in apassive predictive role or an active`closed-loop' ventilation manage-ment system are potential roles forphysiological simulation.

Key words Physiologicalsimulation × Respiratory model ×Mechanical ventilation

Intensive Care Med (1999) 25: 839±842Ó Springer-Verlag 1999 BRIEF REPORT

Patients' pathologies included multiple trauma, respiratory failureand systemic inflammatory response syndrome. Changes weremade in the fractional inspired oxygen (FIO2) in all cases and inminute volume via respiratory rate or tidal volume in some cases.Those patients whose positive end expiratory pressure (PEEP)was adjusted were excluded, as modelling of PEEP by the NPS iscurrently empirical and is not validated. We also excluded patientswho had unstable cardiorespiratory physiology (i. e. those patientswhose condition may have changed during the interval betweenventilator adjustment and data acquisition). Each data set includ-ed: (1) arterial blood gas analysis performed using a regularly cali-brated 278 Blood-Gas System (Chiron Diagnostics, Halstead,UK); (2) respiratory rate and expired tidal and minute volumesmeasured by the integral flowmeter of the Dräger II (Dräger Lim-ited, Hemel Hempstead, UK) ventilator; (3) haemoglobin concen-tration measured daily in the hospital laboratory; (4) temperaturemeasured using a calibrated axillary skin probe.

Following alignment to each patient, the simulator was subject-ed to the same changes in ventilator settings as the patient and theresulting (steady-state) PaO2, PaCO2 and arterial pH were record-ed. Six consultant (specialist) ICU physicians were then providedwith the patients' pre-intervention data sets and details of thechanges in ventilator settings. All of the physicians were experi-enced in managing mechanical ventilation, and managed ventilat-ed patients on a daily basis. Each provided a prediction of thesteady-state PaO2, PaCO2 and arterial pH following intervention.They were asked to perform this estimation as they would in a clin-ical setting.

The predicted size of change in PaO2, PaCO2 and arterial pHwere considered in order to evaluate the accuracy of the NPS and

the physicians. Bias in predicting the size of change of each of thesevariables was calculated as: mean (predicted ± measured) size ofchange. Ninety five per cent limits of agreement (LA95 %) betweenthe predicted and the measured size of change in each variablewere calculated as 95 % confidence intervals (CI) of the bias [9];95% CI were calculated as mean � 1.95 � standard deviation.Data recording, analysis and charting were performed using Mi-crosoft Excel (version: Office 97).

Results

Twenty-five data sets were recorded. No data sets wereexcluded from the study. FIO2 was adjusted in all 25 pa-tients and minute volume was adjusted in 13 of these pa-tients. Mean magnitude of change in FIO2 was 0.15 (SD0.06) and mean magnitude of change in minute volumewas 700 (SD 1000) ml. Mean absolute changes in pa-tients following intervention were: PaO2 5.95 (SD 3.23)kPa, PaCO2 0.62 (SD 0.67) kPa and arterial pH 0.041(SD 0.032). Figure 1 shows the bias and 95 % CIs in pre-diction of size of change of PaO2, PaCO2 and arterial pHby the NPS and the physicians [1±6].

Discussion

We analysed magnitude of change in blood gas variablesrather than direction of change. It was assumed that inthe majority of cases, prediction of the direction ofchange was relatively trivial and that consideration ofsize of change allowed more sensitive analysis of accura-cy [6]. The NPS was more accurate and consistent thanany of the six physicians in predicting magnitude ofchange of PaO2 and arterial pH. The simulator overesti-mated the magnitude of PaCO2 change and had a larger

840

Fig.1 Bias in predicting size of change in PaO2 filled circles, PaCO2open squares and arterial pH open diamonds for the simulator NPSand the physicians 1±6. Error bars represent LA95% between pre-dicted and measured size of change. Negative bias represents anunderestimation of the size of change and positive bias representsan overestimation in size of change. Thus, it may be seen that theNPS slightly overestimated the size of change of all three variablesfollowing intervention

bias than four of the physicians, but was more consistent(narrower LA95 %) than all but one of the physicians.The simulator tended to overestimate slightly the mag-nitude of change of all variables. This could be due tocoincidental changes in patients' physiological vari-ables, inaccuracies in the physiological modelling or er-rors in measurement of those variables. However, thereason was most likely to have been an inadequateequilibration time between intervention and sampling.We chose 10 min as our equilibration time as we feltthat any longer may result in inaccuracies due to coinci-dental physiological changes in this unstable group ofpatients. In addition, nursing and medical interventionwas discouraged between changes in ventilator settingsand data acquisition, and it was felt that this could notbe prolonged for more than 10 min. Previous studieshave used longer equilibration periods, but have ac-quired data retrospectively [10].

The physicians showed individual patterns of inaccu-racy in prediction, reflecting their erroneous applicationof respiratory pathophysiology. An example of thesesystematic misinterpretations is given in Fig. 2, whichshows the predictions of changes in PaO2 for one of thephysicians and the NPS against the observed changes.The simulator showed consistency in prediction throughthe whole range of values measured, while the physi-cians demonstrated a variety of systematic errors. Thishas implications for patient care since physicians may

erroneously adjust mechanical ventilation in an attemptto attain optimal arterial gas tensions and pH. The NPSmay prove to be a useful tool in the management of me-chanical ventilation and may be incorporated into me-chanical ventilators to allow a `trial of ventilator adjust-ment' or may eventually form part of a `closed-loop'ventilation management system.

Appendix

The Nottingham Physiology SimulatorThe NPS comprises a set of advanced, original physiological mod-els. The models use iterative refinement to calculate the dynamicequilibria found in vivo. The respiratory models include serial andparallel deadspaces and variable ventilation-perfusion matching.Alveolar contents are continuously determined by the dynamic re-lationship between the pulmonary blood, alveolar contents andgases inhaled into the alveoli. Blood carriage of oxygen is depen-dent upon temperature, pH and CO2 tension. Partial pressures ofnitrogen and water are explicitly modelled.

Aligning the simulation to each patientThe haemoglobin concentration, FIO2, temperature, tidal volume,respiratory rate and plasma base excess are entered into the simu-lation. The alignment algorithm now matches the PaO2 andPaCO2 in the NPS with the patient's by simultaneously adjustingthe oxygen consumption, shunt fraction, respiratory quotient andalveolar deadspace fraction. This solution does not necessarily pro-vide correct values for oxygen consumption, shunt, respiratoryquotient and deadspace, but is a possible solution. The simulatornext allocates each of the above factors a probability of existing inisolation from a set of confidence intervals and calculates the com-pound probability for the entire solution. The variables are thenfurther adjusted to maximise this compound probability whilemaintaining a possible physiological solution for the data set.

841

Fig.2 Predicted vs. measured change in PaO2 for the NPS and oneof the specialist physicians. The dashed line represents the line ofidentity upon which perfectly predicted PaO2 values would lie.This physician tended to underestimate the resulting PaO2 whenFIO2 was increased

842

References

1. Mathru M, Rao TL, Venus B (1983)Ventilator-induced barotrauma in con-trolled mechanical ventilation versusintermittent mandatory ventilation.Crit Care Med 11: 359±361

2. Pollack MM, Fields AI, Holbrook PR(1979) Pneumothorax and pneumome-diastinum during pediatric mechanicalventilation. Crit Care Med 7: 536±539

3. Fisher AB (1980) Oxygen therapy. Sideeffects and toxicity. Am Rev RespirDis 122: 61±69

4. Jenkinson SG (1993) Oxygen toxicity.New Horiz 1: 504±511

5. Hinds CJ, Roberts MJ, Ingram D, Dick-inson CJ (1984) Computer simulationto predict patient responses to altera-tions in the ventilation regime. Inten-sive Care Med 10: 13±22

6. Hardman JG, Bedforth NM, AhmedAB, Mahajan RP, Aitkenhead AR(1998) A physiology simulator: valida-tion of its respiratory components andits ability to predict the patient's re-sponse to changes in mechanical venti-lation. Br J Anaesth 81: 327±332

7. Hardman JG, Bedforth NM, MahajanRP, Aitkenhead AR (1998) Predictingthe response to adjustment of mechani-cal ventilation using a physiologicalsimulator (abstract). Br J Anaesth 80:549P

8. Hardman JG, Bedforth NM, MahajanRP, Aitkenhead AR (1998) Estimatingpulmonary shunt fraction in hypercap-nic patients using a physiology simula-tor. British Journal of Anaesthesia80(1): 77. (Abstract)

9. Bland JM, Altman DG (1986) Statisti-cal method for assessing agreement be-tween two methods of clinical measure-ment. Lancet I:307±310

10. Rutledge GW, Thomsen GE, Farr BR,Tovar MA, Polaschek JX, Beinlich IA,Sheiner LB, Fagan LM (1993) The de-sign and implementation of a ventila-tor-management advisor. Art IntellMed 5: 67±82