 Research
 Open Access
 Published:
Iterative integral parameter identification of a respiratory mechanics model
BioMedical Engineering OnLinevolume 11, Article number: 38 (2012)
Abstract
Background
Patientspecific respiratory mechanics models can support the evaluation of optimal lung protective ventilator settings during ventilation therapy. Clinical application requires that the individual’s model parameter values must be identified with information available at the bedside. Multiple linear regression or gradientbased parameter identification methods are highly sensitive to noise and initial parameter estimates. Thus, they are difficult to apply at the bedside to support therapeutic decisions.
Methods
An iterative integral parameter identification method is applied to a second order respiratory mechanics model. The method is compared to the commonly used regression methods and errormapping approaches using simulated and clinical data. The clinical potential of the method was evaluated on data from 13 Acute Respiratory Distress Syndrome (ARDS) patients.
Results
The iterative integral method converged to error minima 350 times faster than the Simplex Search Method using simulation data sets and 50 times faster using clinical data sets. Established regression methods reported erroneous results due to sensitivity to noise. In contrast, the iterative integral method was effective independent of initial parameter estimations, and converged successfully in each case tested.
Conclusion
These investigations reveal that the iterative integral method is beneficial with respect to computing time, operator independence and robustness, and thus applicable at the bedside for this clinical application.
Introduction
Mechanical ventilation (MV) therapy in intensive care units (ICU) carries the risk of severe additional complications for the patient due to nonoptimal ventilator settings [1]. To reduce patient risk, optimal patientspecific ventilator settings must be found. Mathematical models of respiratory mechanics effectively predict the outcome of specific ventilator settings and thus support clinical evaluation of lung protective settings [2]. Prediction quality depends on how well model parameters correspond to the true patient properties and on the accuracy of the model representing the individual’s lung physiology. To assure clinical applicability at the bedside, especially in closed loop settings [2, 3], reliable and immediate parameter identification is required. However, information for analyzing respiratory mechanics directly at the bedside is generally restricted to measurements of airway pressure and flow rate. Therefore, any model must enable easy and fast parameterization in terms of the available clinical data. Thus, models should capture all necessary dynamics and be as simple as possible. Hence, lumped parameter models are a common approach for optimizing ventilator settings [4].
Typical parameter identification methods minimize the least square error (LSE) between measured samples and model simulations. For single compartment models of respiratory mechanics, multiple linear regression is an established straightforward approach [5, 6]. Multiple linear regression is also applicable to higher order linear models, but the quality of results can be distorted by noise [6]. Alternative parameter identification methods for higherorder and nonlinear models are iterative errormapping methods [7–11], such as the SimplexSearch Method [12] or LevenbergMarquardt Algorithm [13, 14]. These methods require initial guesses for the model parameters and iterate towards LSE values by approaching the minimum on the error surface. However, with an increasing number of degrees of freedom, several parameter constellations can appear as additional possible solutions (local minima) [8]. Hence, such methods can report nonoptimal parameter values and are highly dependent on initial parameter estimates. To reduce the probability of convergence to local minima, various global search strategies, such as random search, simulated annealing or genetic algorithms can be used [15]. However, these methods are time consuming and the quality of the solution is sensitive to the parameterization of the algorithm [16].
This research presents an iterative integral method (IIM) [17, 18] for respiratory mechanics parameter identification, and compares the performance against established parameter identification methods. The method was originally developed for parameter identification of glucoseinsulin models [19]. In the context of the glucoseinsulin models, the IIM is comparatively simple to apply, requires comparatively minimal computing time and does not require the estimation of initial values [17, 19]. The IIM is tested with simulation data to demonstrate its robustness and efficiency, and clinical study data is used to demonstrate the clinical potential.
Model & methodology
Viscoelastic model
Viscoelastic Models (VEM) of respiratory mechanics are established models and assume that the tissues comprising the walls of the alveolar compartment are viscoelastic, rather than simply elastic [20]. The VEM used in this research is an established model of respiratory mechanics whose applicability with respect to healthy and lungs with Acute Respiratory Distress Syndrome (ARDS) has been demonstrated [21, 22]. Ganzert et al. [10] showed the different trends of the VEM parameters in healthy and ARDS subjects, which might be used as indicator for ARDS development.
The analogous electrical circuit for the VEM is shown in Figure 1 and the mathematical description is presented in statespace representation in Eq. (1):
The flow rate $\dot{V}$ [mL/s] defines the model input and the airway pressure p _{aw} [cmH_{2}O] defines the model output. Airway resistance is described by R _{1} [cmH_{2}O·s/mL] and the static lung compliance by C _{ 1 } [mL/cmH_{2}O]. R _{ 2 } [cmH_{2}O·s/mL] and C _{2} [mL/cmH_{2}O] account for the viscoelasticity described by the viscoelastic time constant R _{ 2 }·C _{ 2 }. p _{ C1 } and p _{ C2 } [cmH_{2}O] are the internal state variables being the pressure in the compliant static and viscoelastic lung compartment.
The four parameters (R _{1}, C _{ 1 }, R _{2}, C _{2}) correspond to the unknown patientspecific parameters:
A fundamental prerequisite for successful parameter identification is apriori structural identifiability of the parametric model [23]. This necessary criterion states that under ideal conditions of noisefree observations and errorfree model structure, the unknown parameters of the model can be uniquely recovered from the measured input–output variables. The VEM is globally identifiable [8] using the DAISY (Differential Algebra for Identifiability Systems) computer program to automatically check for apriori structural identifiability [23].
Data
First, the parameter identification methods were compared insilico using synthetic data sets generated by simulating parameterized VEMs. Second, the methods were tested using clinical data to assess their clinical applicability.
Simulation data
A flow profile corresponding to the VolumeControlledVentilation (VCV) mode was applied to parameterized VEMs. Inflation was simulated by a constant flow of 500 mL/s over 1 s, followed by an endinspiratory pause (EIP) of 4 s corresponding to the closure of the expiratory valve as observed in clinical data [24].
100 different sets of model parameters were generated for a MonteCarlo Analysis. The model parameters were randomly selected from an evenly distributed range between 0.5 and 1.5 times a realistic physiological parameter set (R _{ 1 } = 0.010 cmH_{2}O·s/mL, C _{ 1 } = 30.00 mL/cmH_{2}O, R _{ 2 } = 0.020 cmH_{2}O·s/mL, C _{ 2 } = 80.00 mL/cmH_{2}O).
After simulating each VEM, 10 different sets of white noise (5% of given pressure sample) were added in order to impose noisy measurement error. An example of the simulated pressure response is shown in Figure 2. It consists of an increase during inspiration and an exponential drop during the EIP.
Clinical data
Measurement sets from N = 13 mechanically ventilated patients were selected from a previous ARDS– Study [10, 25], where SCASSManeuvers (Static Compliance Automated Single Step) were performed. The SCASSManeuver consists of airway occlusions within the inspiration phase of a breathing cycle that are initiated when a randomized inspiration volume is reached. The measurement set consisted of flow rate ($\dot{V}$) and airway opening pressure (p _{ aw }) signals sampled at 125 Hz. For each patient, two breathing cycles were extracted. During occlusion, the airway pressure reached a quasistatic equilibrium exponentially.
The study was approved by the local ethics committees of the participating university hospitals. Informed consent was obtained from patients or their legally authorized representative. The detailed description of the experimental setup is shown in [25].
Parameter identification
Five parameter identification methods were used to identify the patientspecific VEM parameters Eq. (2) using the simulation and clinical data. These methods also provide references for the iterative integral method. To compare computational efficiency, all methods were tested on the same Desktop PC (Intel Core 2 Duo, 2.80 GHz), where computation time was measured.
Multiple linear regression (MLR)
Given a linear model, MLR is an obvious approach for parameter identification [5, 6]. MLR requires the input–output relation of the model, eliminating the state variables. This relation is derived from the state space description of the VEM by calculating the derivative of the output equation from Eq. (1) and inserting the definition of the state variables:
To eliminate p _{ C2 } in Eq. (3), the output of Eq. (1) is rewritten in terms of p _{ C2 }:
Eq. (4) is inserted in Eq. (3) leading to the input–output relation of the model:
Effectively, Eq. (6) rearranges the statespace equation Eq. (1) to place the model variables as functions of measured values (p _{ aw } and $\dot{V}$) and their derivatives or integrals.
MLR requires a sum of independent variables scaled by a multiplicative factor. Therefore the coefficients of Eq. (6) can be represented by new variables:
For parameter identification according to the LeastSquareError (LSE) principle, Eq. (7) was arranged in matrix form:
Values of AD are identified applying the MoorePenrose pseudodivision for overdefined matrix equations. Patientspecific R _{ 1 } , C _{ 1 } , R _{ 2 } , C _{ 2 } values are then uniquely derived from the identified values of AD.
Integral Method (IM)
The IM is similar to MLR in terms of identification steps. However, integrals are used to improve robustness to noise [19]. Hence, Eq. (5) was integrated assuming p _{ aw }(0) = 0:
The coefficients of Eq. (9) are represented by new variables:
Yielding:
Incorporating Eq. (11) into an overdefined matrix system yields:
Patientspecific parameters R _{ 1 } , C _{ 1 } , R _{ 2 } , C _{ 2 } are uniquely regained using the identified values of AD:
Iterative Integral Method (IIM)
The IIM uses much the same method as the IM. However, the identified AD values are used to resimulate p _{ aw } according to Eq. (11). Resimulated p _{ aw } is then used to update the lefthandside (LHS) integrals of Eq. (12). Subsequently, Eq. (12) is solved again yielding updated values of AD. This process is repeated until convergence. The sum of squared error (SSE) between simulated and measured p _{ aw } was calculated after every iteration. The relative termination tolerance (TolFUN) for changes in the function value (SSE) was set to 10^{4} This value is in accordance to the default setting of the proprietary gradientbased search methods in MATLAB.
Proprietary LevenbergMarquardt Algorithm (LMA) and Simplex Search Method (SSM)
GradientBased Methods such as the LevenbergMarquardt Algorithm (LMA) (MALTAB command: lsqnonlin) and the SimplexSearch Method (SSM) (MALTAB command: fminsearch) were applied as a references to the IIM using following default convergence criteria: terminate the optimization if the relative change in parameters is smaller than 10^{4} (MATLAB: TolX value) or when the relative change in the SSE value is smaller than 10^{4} (MATLAB: TolFUN value).
GradientBased Methods require initial estimates for the corresponding parameters. Initial estimates close to the global minimum increase the probability of successful parameter identification.
For the case of identifying simulated data, the parameter values that were used for simulation are known and were used as initial values to provide the fastest, easiest solution, and a conservative comparison.
For the case of clinical data, patient outcomes are initially unknown. Thus, convenient estimates for initial values are important. To decrease the potentially deleterious influence of poorly estimated initial values for gradientbased methods, a hierarchical method was developed by Schranz et al. [8]. This approach uses identification of simple respiratory models to provide prior knowledge. It is preferable for the simpler models to be identified with linear regression to avoid the influence of initial values. The identified parameter values are then used as initial values to support the identification of more complex models. In the VEM case, the 1^{st} order model (FOM) of respiratory mechanics (Figure 3) and Equation (14) was first identified by MLR leading to estimates of airway resistance (R) and respiratory compliance (C).
Additionally, the time constant τ of the exponential pressure drop during the zeroflow phase (Figure 2) was estimated by extracting the corresponding pressure interval fitted by an exponential function. Given an errorfree model structure, this estimated time constant corresponds to the viscoelastic time constant R _{ 2 } ·C _{ 2 }. With the information of R, C and τ, initial values can be provided for gradientbased parameter identification of the VEM. This method improved the robustness of the subsequent parameter identification significantly by reducing the potential for poorly estimated initial values to result in local minima identification [8]. It is thus a more effective and conservative method comparison for the IIM.
Results
Parameter identification using simulation data
The identified parameter values from each parameter identification method were equal in the noise free synthetic data set and were effectively equal to the simulation values (R = 1.000). However, adding noise to the simulation data causes the true global minimum to slightly shift away from the original parameter values. In this case, the solutions reported by the SSM were used to define the new error minimum. The MonteCarlo analysis revealed that the IIM, IM, SSM and LMA reported the same solution error minimum in every data set. The MLR didn’t locate a reasonable error minimum in any single case. The resulting parameter values of the MLR deviate by up to 100% from the values at the reported error minima. This outcome led to a median SSE 380% larger than the SSE of the error minima. The IIM always reached convergence after its first iteration and was thus equivalent to the IM in this case. The required median [interquartile] computing time of the IM and IIM was 0.8 [0.70.9] ms. In comparison, the gradientbased methods require 108.3 [101.6115.7] ms for the SSM and 15.1 [15.0 – 15.5] ms for LMA.
Parameter identification using clinical data
All results using the clinical data are summarized in Table 1. Within all clinical data sets, the SSM converged to a minimal SSE with physiologically plausible parameter values. These values are considered as error minimum. The SSM required an average of 251 ms per data set. The LMA found the same minima in 17 data sets, but located local nonphysiological minima with partly negative parameter values and higher SSE values in 9 out of 26 (35%) data sets. MLR identification resulted in nonphysiological parameter values and significantly higher SSE values in every (100%) data set. IM identification reported nonphysiological parameter values that were either negative or significantly higher than the error minimum in 8 of 26 (31%) data sets with higher SSE values than the SSM. Parameter identification using the IIM resulted in similar minima as the SSM in every data set.
Table 2 shows the IIM convergence in an example data set where the IM reported erroneous results. Starting from the erroneous IM result (0^{th} iteration), the SSE and nonphysiological parameter values converge step by step towards the error minimum approaching physiologically plausible parameter values.
The median required computing time in Table 1 for the IIM equals 5.02 ms reaching the minimum error faster than the SSM by a factor of 50. Example pressure responses for parameter sets identified by IIM, IM and MLR are shown together with the residuals in Figure 4. The residuals of the pressure response produced by IM parameter values are higher than the response of IIM parameter values. The pressure responses of the parameter sets identified by SSM and LMA are congruent with the solution of the IIM. Therefore these plots are neglected in Figure 4.
Figure 5 shows the correlation plots of the reported IIM and SSM parameter values and confirms that the same solutions were reached in each data set. As the termination of the method is determined by convergence tolerances, minimal deviations occur in each parameter, leading to correlations lower than 1.000, ranging from R = 0.986 – 1.000. The correlation for the MLR, IM and LMA to the SSM are R = −0.209 –0.992, R = −0.448 – 0.999, R = 0.158 – 0.999, respectively.
Discussion
The noisefree insilico analysis showed that all identification methods where capable of identifying the parent parameter values via the simulated model response. This result confirms that the model is structurally identifiable [8, 23], and the identification methods were accurately implemented. When limited white noise was added to the simulated profile all tested methods, except MLR successfully located a similar error minimum with identified parameter values close to their parent values. MLR resulted in significantly higher SSE values in every single data set and parameter values that were highly removed from their parent values. These erroneous results are caused by the noiseamplifying effects of derivative terms in Equation (6). The differentiation terms in Equation (6) are due to the secondorder characteristics of the underlying twocompartment model. However, for some simpler models (e.g. linear and nonlinear onecompartment models) MLR can be a reasonable and legitimate method for parameter identification [5, 8]. However, this analysis indicates that MLR should not be applied for parameter identification of this VEM or such other higherorder models.
While the IM utilizes much the same methodology as the MLR, the outcomes of the noisy insilico analysis and the clinical analysis show the merit of integrating the governing equation. The integrals in the IM lowpass filter the measured signals, yielding significant improvement in robustness over MLR for noisy simulation data. However, the IM located some nonphysiological parameter values in the clinical data. In total, there were negative parameter values in 8 of 26 (31%) data sets. The erroneous parameter constellations resulted in higher residuals than reported by SSM and IIM methods, but still produced reasonable simulation results. These unacceptable parameter values results are in accordance with the findings of Sato et al [6], who found improvement in integralbased VEM identification over MLR, but still obtained negative parameter values. The IM failure to consistently produce accurate parameter values might be due to unmodeled effects in the p _{ aw } (t) term and the subsequent amplification of error via the comparatively highly parameterized model. Less parameterized models have been robust to such estimation of the simulated profile and successfully identified with the IM [19, 26, 27].
In contrast to the IM, the IIM enabled successful parameter identification in every case tested. Table 2 highlights a case wherein the IM (effectively the first cycle of IIM) yielded a highly unphysiological value for C _{ 2 }. In this case, the IIM converged to a lower error state with more plausible physiological values. By repeatedly evaluating the matrix equation Eq. (12) the p _{ aw } (t) term in the left hand side is updated, whereas the p _{ aw } (t) on the right hand side remains the measured data. Thus, the linear least square error minimization process optimizes the simulated p _{ aw } (t) against the measured values directly. Figure 5 shows that the IIM and the SSM located the same parameter values and Table 1 shows that the methods found the lowest overall error values.
The LMA is a popular parameter identification method for linear and nonlinear problems [9, 28, 29]. LMA parameter identification in clinical data was 3.5 times faster than the SSM due to its higher order, but still 14 times slower than the IIM. However, the LMA appears more sensitive to initial values than the SSM. With the same hierarchically derived initial values, the LMA reported local minima with comparatively high SSE values in 9 out of 26 (35%) data sets. Based on this study, the LMA with hierarchical support is not suitable for reliable straightforward VEM parameter identification.
The IIM and SSM located very similar minimal error in every case. In contrast, MLR consistently failed in the presence of measurement error and noise. The IM was unable to report the error minima located by IIM and SSM, perhaps due to unmodeled effects in p _{ aw }. LMA occasionally converged successfully, but located local minima in a significant number of cases. Hence, only the IIM and SSM could be considered as reliable methods for VEM parameter identification. The robustness of the SSM was dependent on the initial values, which were provided here by a hierarchical method. In contrast, the IIM was independent of initial values, which is a significant reduction in complexity and is thus more robust in application.
Limitations
All models are representations of a true system and thus all models have limitations. In this case, the underlying VEM is a basic model and is limited in its abilities to describe all effects in respiratory mechanics, in large part to capture fundamental physiology and also to ensure it is identifiable with readily available data or simple maneuvers. Pathophysiological effects that are not captured by the VEM include for example alveolar recruitment and overdistension, flow limitation, and edema among leading issues that can affect such models and experiments. Assuming these unmodeled effects are not present in the data, a perfect model agreement could be guaranteed, given a perfect experiment.
However, additional uncontrolled effects are always present in clinical data due to measurement noise and the clinical impact of these devices on physiological responses. Secondly, inadequate experimental handling e.g. spontaneous muscle activities caused by insufficient sedation interferes with the assumed experimental hypotheses and model assumptions that neglect this activity. With this variety of unmodeled effects in the data, the IM can report inaccurate results that could be corrected by applying the IIM.
For example, assuming a hypothetical impulse of pressure in the measured data at t = t _{ 2 } due to any of the above described reasons, then the integral of the pressure will be affected from t _{ 2 } to t _{ end } in Eq 8. Hence, the IM will have the aberration in both sides of the equation and the model functions that are most prominent in the period t _{ 2 } will be negatively affected. In contrast, the IIM uses a resimulated model response and thus, the impulse gets smoothed in the left hand side of Eq 8 over one or more iterations. The error induced by the impulse at t = t _{ 2 } is present in the model simulation contained in the left hand side of Eq 8 from t = 0 to t _{ end }. The error is not limited to a particular period of the response and thus the error does not exaggerate any particular model function. With ongoing iterations the influence of the impulse is reduced, and the results of the IIM approach the solution of the gradientbased methods. Hence, the IIM offers builtin modelbased compensation for such errors within its framework, as the iterations enforce the model dynamics onto the data mitigating the impact of unmodeled effects. These effects are presumed small, although it should be noted that this presumes the model itself captures the fundamental mechanics accurately.
Small deviations within the reported parameter values of both methods were found (Figure 5). These deviations are well within clinical tolerances and were unlikely to significantly alter therapeutic choices. The lowest correlation (R = 0.986) of the IIM parameters to the SSM parameters were found in R _{ 2 }, which is, together with C _{ 2 } , predominantly sensitive to the relaxation phase [30]. This phase is mostlikely to be disturbed by cardiogenic oscillations. Additionally, in cases with no distinct exponential pressure drops during EIP, information content for R _{ 2 } , C _{ 2 } identification is low, decreasing the convexity in the error plane. Low convexity causes an early termination of the tested identification algorithms leading to minor deviation in the corresponding parameters. Changes in tolerances used would also improve these small differences but are with a variation of less than 1% utilizing a TolFun value being 10 times smaller, relatively small.
Technical and clinical impact
The IIM is potentially applicable to other linear respiratory mechanics models of higher order, such as the Inhomogeneity Model [31], as this model poses a transfer function of the same form as the VEM [32]. Furthermore, the advantage of the IIM of being independent on initial values is beneficial for the principle of hierarchical approaches of parameter identification. A robust identification process of more advanced models like the VEM avoids gradientbased methods within this hierarchical level and shifts these more susceptible methods further down in the model hierarchy. Thus, an implementation of IIM would lead to increased robustness and efficiency in hierarchical approaches of more extensive hierarchical model structures [33].
Applying the IIM to the VEM allows onlinemonitoring the trends of patientspecific viscoelasticity at the bedside, and thus supports controlling the mechanical ventilator to maintain the viscoelastic properties in a normal range. Secondly, using a frequency dependent model supports the adjustment of ventilation frequency at the lowest impedance. Further clinical studies are required to develop a control scheme based on viscoelastic properties of the individual patient’s lung.
Conclusion
Compared to gradientbased methods, the IIM provides robust parameter identification and operates without the initial value problem. Therefore it can offer more operator independence in model parameter identification.
The VEM investigations confirm that parameter identification via the IIM is also superior to errormapping methods with respect to efficiency and robustness. The IIM takes advantage of the linear mathematical structure of the model and offers fast computing time and maximal robustness. These features offer potential for the IIM to be implemented in an online tool at the bedside for ventilation management.
Abbreviations
 ARDS:

Acute Respiratory Distress Syndrome
 C:

Compliance in FOM
 C1:

Static compliance in VEM
 C2:

Viscoelastic compliance in VEM
 EIP:

Endinspiratory pause
 FOM:

1st Order Model
 ICU:

Intensive Care Unit
 IIM:

Iterative Integral Method
 IM:

Integral Method
 LMA:

LevenbergMarquardt Algorithm
 LSE:

Least Square Error
 MLR:

Multiple linear regression
 MV:

Mechanical Ventilation
 paw:

Airway pressure
 pC1:

Pressure in C1
 pC2:

Pressure in C2
 R:

Resistance in FOM
 R1:

Airway resistance in VEM
 R2:

Viscoelastic resistance in VEM
 SSE:

Sum of squared errors
 SSM:

SimplexSearch Method
 TolFun:

Terminating Function Tolerance
 TolX:

Terminating Parameter Tolerance
 VCV:

VolumeControlledVentilation
 VEM:

Viscoelastic Model
 XVEM:

Unknown patientspecific VEM parameter
 τ:

Time constant
 V˙:

Flow rate.
References
 1.
Slutsky AS: Lung injury caused by mechanical ventilation. Chest 1999, 116: 9–15. 10.1378/chest.116.suppl_1.9S
 2.
Lozano S, Möller K, Brendle A, Gottlieb D, Schumann S, Stahl CA, Guttmann J: AUTOPILOTBT: a system for knowledge and model based mechanical ventilation. Technol Health Care 2008, 16: 1–11.
 3.
Rees SE, Allerød C, Murley D, Zhao Y, Smith BW, Kjaergaard S, Thorgaard P, Andreassen S: Using physiological models and decision theory for selecting appropriate ventilator settings. J Clin Monit Comput 2006, 20: 421–429. 10.1007/s1087700690495
 4.
Hann CE, Chase JG, Desaive T, Froissart CB, Revie J, Stevenson D, Lambermont B, Ghuysen A, Kolh P, Shaw GM: Unique parameter identification for cardiac diagnosis in critical care using minimal data sets. Comput Methods Prog Biomed 2010, 99: 75–87. 10.1016/j.cmpb.2010.01.002
 5.
Polak AG: Analysis of multiple linear regression algorithms used for respiratory mechanics monitoring during artificial ventilation. Comput Methods Prog Biomed 2011, 101: 126–134. 10.1016/j.cmpb.2010.08.001
 6.
Sato J, Davey BL, Shardonofsky F, Bates JH: Lowfrequency respiratory system resistance in the normal dog during mechanical ventilation. J Appl Physiol 1991, 70: 1536–1543.
 7.
Sundaresan A, Yuta T, Hann CE, Chase JG, Shaw GM: A Minimal Model of Lung Mechanics and Model based Markers for Optimizing Ventilator Treatment in ARDS Patients. Comput Methods Programs Biomed 2009, 95: 166–180. 10.1016/j.cmpb.2009.02.008
 8.
Schranz C, Knöbel C, Kretschmer J, Zhao Z, Möller K: Hierarchical Parameter Identification in Models of Respiratory Mechanics. IEEE Trans Biomed Eng 2011, 58: 3234–3241.
 9.
Diong B, Nazeran H, Nava P, Goldman M: Modeling Human Respiratory Impedance. IEEE Eng Med Biol Mag 2007, 26: 48–55.
 10.
Ganzert S, Möller K, Steinmann D, Schumann S, Guttmann J: Pressuredependent stress relaxation in acute respiratory distress syndrome and healthy lungs: an investigation based on a viscoelastic model. Crit Care 2009, 13: R199. 10.1186/cc8203
 11.
Khoo MC, Yamashiro SM, Yamashiro P: Minimization of lung pressure swings during highfrequency ventilation: a model. J Appl Physiol 1989, 67: 993–1000.
 12.
Lagarias JC, Reeds JA, Wright MH, Wright PE: Convergence Properties of the NelderMead Simplex Method in Low Dimensions. SIAM J Optim 1998, 9: 112–147. 10.1137/S1052623496303470
 13.
Levenberg K: A Method for the Solution of Certain Problems in Least Squares. Quart Appl Math 1944, 2: 164–168.
 14.
Marquardt DW: An Algorithm for LeastSquares Estimation of Nonlinear Parameters. SIAM J Optim Appl Math 1963, 11: 431–441. 10.1137/0111030
 15.
Ljung L: System Identification  Theory for the User. PrenticeHall, Englewood Cliffs, New Jersey; 1999.
 16.
Ingber L: Simulated Annealing: Practice versus Theory. Mathl Comput Modelling 1993, 18: 29–57.
 17.
Docherty PD, Chase JG, Lotz T, Hann CE, Shaw GM, Berkeley JE, Mann JI, McAuley K: DISTq: An Iterative Analysis of Glucose Data for LowCost, RealTime and Accurate Estimation of Insulin Sensitivity. Open Med Inform J 2009, 3: 65–76. 10.2174/1874431100903010065
 18.
Docherty PD, Chase JG, David T: Characterisation of the iterative integral parameter identification method. Med Biol Eng Comput 2012, 50: 127–134. 10.1007/s115170110851y
 19.
Hann CE, Chase JG, Lin J, Lotz T, Doran CV, Shaw GM: Integralbased parameter identification for longterm dynamic verification of a glucoseinsulin system model. Comput Methods Prog Biomed 2005, 77: 259–270. 10.1016/j.cmpb.2004.10.006
 20.
Lutchen KR, Costa KD: Physiological interpretations based on lumped element models fit to respiratory impedance data: use of forwardinverse modeling. IEEE Trans Biomed Eng 1990, 37: 1076–1086. 10.1109/10.61033
 21.
Jonson B, Beydon L, Brauer K, Mansson C, Valind S, Grytzell H: Mechanics of respiratory system in healthy anesthetized humans with emphasis on viscoelastic properties. J Appl Physiol 1993, 75: 132–140.
 22.
Beydon L, Svantesson C, Brauer K, Lemaire F, Jonson B: Respiratory mechanics in patients ventilated for critical lung disease. Eur Respir J 1996, 9: 262–273. 10.1183/09031936.96.09020262
 23.
Bellu G, Saccomani MP, Audoly S, D'Angio L: DAISY: A new software tool to test global identifiability of biological and physiological systems. Comput Methods Prog Biomed 2007, 88: 52–61. 10.1016/j.cmpb.2007.07.002
 24.
Sydow M, Burchardi H, Zinserling J, Ische H, Crozier TA, Weyland W: Improved determination of static compliance by automated single volume steps in ventilated patients. Intensive Care Med 1991, 17: 108–114. 10.1007/BF01691433
 25.
Stahl CA, Möller K, Schumann S, Kuhlen R, Sydow M, Putensen C, Guttmann J: Dynamic versus static respiratory mechanics in acute lung injury and acute respiratory distress syndrome. Crit Care Med 2006, 34: 2090–2098. 10.1097/01.CCM.0000227220.67613.0D
 26.
McAuley KA, Berkeley JE, Docherty PD, Lotz TF, Te Morenga LA, Shaw GM, Williams SM, Chase JG, Mann JI: The dynamic insulin sensitivity and secretion test–a novel measure of insulin sensitivity. Metabolism 2011, 60: 1748–1756. 10.1016/j.metabol.2011.05.009
 27.
Lotz TF, Chase JG, McAuley KA, Shaw GM, Docherty PD, Berkeley JE, Williams SM, Hann CE, Mann JI: Design and clinical pilot testing of the modelbased dynamic insulin sensitivity and secretion test (DISST). J Diabetes Sci Technol 2010, 4: 1408–1423.
 28.
Raue A, Kreutz C, Maiwald T, Bachmann J, Schilling M, Klingmuller U, Timmer J: Structural and practical identifiability analysis of partially observed dynamical models by exploiting the profile likelihood. Bioinformatics 2009, 25: 1923–1929. 10.1093/bioinformatics/btp358
 29.
Liu CH, Niranjan SC, Clark JW, San KY, Zwischenberger JB, Bidani A: Airway mechanics, gas exchange, and blood flow in a nonlinear model of the normal human lung. J Appl Physiol 1998, 84: 1447–1469.
 30.
Thomaseth K, Cobelli C: Generalized sensitivity functions in physiological system identification. Ann Biomed Eng 1999, 27: 607–616.
 31.
Bates JH: Lung Mechanics: An Inverse Modeling Approach. 1st edition. Cambridge University Press, New York, USA; 2009.
 32.
Avanzolini G, Barbini P: Comments on "Estimating respiratory mechanical parameters in parallel compartment models". IEEE Trans Biomed Eng 1982, 29: 772–774.
 33.
Schranz C, Möller K: Modelbased Quantification of Pressure and Time Depending Effects in ARDS Patients. In 5th International Conference on Bioinformatics and Biomedical Engineering (iCBBE). IEEE, Wuhan, China; 2011.
Acknowledgement
The authors express their gratitude to the McREM Study Group and Dräger Medical for providing the data for this retrospective study. This research was supported by the Deutsche Forschungsgemeinschaft DFG under Grant MO 579/12 PAR (Protective Artificial Respiration) and the German Academic Exchange Service (DAAD).
Author information
Additional information
Competing interest
The authors declare that they have no competing interests.
Authors’ contributions
CS, PDD, YSC, JGC developed the methods for this particular model and created the experimental setups. JGC and KM had input to analysis of the results. All authors had input in writing and revising the manuscript.
Authors’ original submitted files for images
Below are the links to the authors’ original submitted files for images.
Rights and permissions
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
About this article
Received
Accepted
Published
DOI
Keywords
 Parameter identification
 Respiratory mechanics
 Viscoelastic model
 Global minimum
 Robustness