| JOURNAL HOME | CME HOME | THIS MONTH | PAST ISSUES | ETOC | COLLECTIONS |
| AUTHORS | REVIEWERS | EDITORIAL BOARD | FEEDBACK | RSS | HELP |
 
|
|
|
||
|
|
Anesth Analg 2003;97:1143-1148 © 2003 International Anesthesia Research Society
|
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
 |
Abstract |
|---|
 Top Abstract IntroductionMethodsResultsDiscussionReferences |
|---|
In this study we evaluated whether a technology based on artificial neural networks (ANN) could estimate the static compliance (CRS) of the respiratory system, even in the absence of an end-inspiratory pause, during continuous mechanical ventilation. A porcine model of acute lung injury was used to provide recordings of different respiratory mechanics conditions. Each recording consisted of 10 or more consecutive breaths in volume-controlled mechanical ventilation, followed by a breath having an end-inspiratory pause used to calculate CRS according to the interrupter technique (IT). The volume-pressure loop of the breath immediately preceding the one with pause was given to the ANN for the training, together with the CRS separately calculated by the IT. The prospective phase consisted of giving only the loops to the trained ANN and comparing the results yielded by it to the compliance separately calculated by the investigators. Determination of measurement agreement between ANN and IT methods showed an error of -0.67 ± 1.52 mL/cm H2O (bias ± SD). We could conclude that ANN, during volume-controlled mechanical ventilation, can extract CRS without needing to stop inspiratory flow.
IMPLICATIONS: We studied the application of artificial neural networks (ANN) to the estimation of respiratory compliance during mechanical ventilation. The study was performed on an animal model of acute lung injury, testing the performance of ANN in both healthy and diseased conditions of the lung.
|
Introduction |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionReferences |
|---|
The elastic properties of the respiratory system are often measured in mechanically ventilated patients to obtain information about lung pathology and to optimize ventilatory strategy. Different methods have been proposed for their assessment during continuous mechanical ventilation. The widely used multilinear fitting (MLF) method (1) can extract respiratory system compliance, assuming that the motion equation of the respiratory gases is composed of linear equations. The pulse method is based on the concept that the compliance of the respiratory system can be determined from the slope of the volume-pressure curve during constant flow inflation (2). Hence, during constant flow-mechanical ventilation, the slope of pressure at airway opening (PAO) plotted versus time during passive inflation is directly related to the elastance of the respiratory system (3,4).
The interrupter technique (IT) derives from the notion that after the inspiratory flow (V’) is stopped, the ratio between the inspired volume and the subsequent pressure change is related mainly to the static components of respiratory mechanics (5). IT, involving the stop of the flow, is considered the gold standard for respiratory mechanics measurement, and for this reason, it has been used in the present study.
Methods based on artificial neural networks (ANNs) enable another approach to the estimation of respiratory mechanics. ANNs are artificial intelligence systems based on the connectionism theory (6). They are universal function approximators (7) and can extract information from different classes of signals after having been trained to perform this specific task by learning from examples.
In the field of ANN application to respiratory mechanics, important contributions have come from Orr and Westenskow (8) on the alarms of anesthesia breathing circuits, Bright et al. (9) on the detection of upper airway obstruction, Léon et al. (10) concerning the detection of esophageal intubation, and Räsänen and Léon (11) on the assessment of lung injury in an animal model.
We demonstrated (12) in an animal model the possibility of assessing the static compliance of the respiratory system (CRS) by providing ANNs with respiratory tracings obtained during an end-inspiratory pause (EIP). In that experimental setup, the extraction of CRS was dependent on the presence of a flow interruption in the tracings to be analyzed by the ANN.
The underlying disease may not allow alterations of breathing pattern to enable the analysis of respiratory mechanics. In these cases, the availability of a tool not requiring such interventions could be beneficial. Considering that information regarding the CRS is present also in the slope of the pressure signal (during constant flow mechanical ventilation), we hypothesized that an EIP should not be required for the assessment of CRS.
Accordingly, the aim of the present experiments was to evaluate whether ANNs can estimate CRS using tracings of PAO and V’ versus time without any intervention of an inspiratory hold maneuver to produce an EIP during continuous mechanical ventilation. This was tested in an animal model during healthy and pathological conditions of the lung.
|
Methods |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionReferences |
|---|
This study is divided into three main phases. The first phase was devoted to the selection of the ANN to be used for extracting CRS from breath not having an EIP. In the second phase, the selected type of ANN was trained and tested using tracings of mechanically ventilated animals presenting different CRS created by the injection of oleic acid (OA). In the third phase of the study, the trained ANN was prospectively evaluated in animals not used for the initial training.
The study was approved by the local IRB for the care of animal subjects. Twenty-four healthy pigs (weight, 29.6 ± 4.6 kg) were included in the study. The group used for training and testing the ANN was composed of 16 animals, whereas the remaining 8 animals comprised the prospective trial group.
After premedication with azaperone, anesthesia was induced by IM administration of atropine, tiletamine-zolazepam, and medetomidin. After oral intubation, total IV anesthesia was started with the administration of ketamine, pancuronium, and fentanyl. Invasive monitoring of central venous, pulmonary artery, and systemic arterial pressure were performed during the experiment; cardiac and urinary outputs were also measured. Fluid replacement strategy was aimed at maintaining a stable systemic arterial pressure. Arterial and mixed venous samples were taken to measure PO2, PCO2, and pH during the various phases of the experiment.
The animals were ventilated using volume-controlled constant-flow mechanical ventilation (VC-MV) (Servo 900 C, Siemens Elema, Solna, Sweden). Tidal volume was adjusted (8–9 mL/kg) to result in normocapnia (35–45 mm Hg) using blood gas samples. Extrinsic positive end-expiratory pressure (PEEPe) was set to 5 cm H2O. Inspiratory fraction of oxygen was 0.5. Inspiratory-to-expiratory ratio was set to 1:2 [s] for a respiratory rate of 20 breaths/min. Acute lung injury was induced by repeated injections of small boluses of OA into the central venous catheter, targeting a total dose of 0.1 mL/kg.
The two sampling ports of a D-Lite connector (Datex Ohmeda, Helsinki, Finland) mounted to the endotracheal tube were connected to a differential pressure transducer (Sensym, SensorTechnics, Pucheim, Germany). At the beginning of each experimental session, the transducer was calibrated for static pressures and for flow measurements. Data were sampled at 200 Hz from the transducer by the Carina 2.4.0 acquisition program (C-O Sjöberg Engineering AB, Upplands-Väsby, Sweden). After real-time collection, traces of V’, PAO, and tidal volume were stored on a personal computer. In both reference and prospective groups of animals, respiratory tracings were recorded at fixed time intervals: after a stabilization period of 60 min after instrumentation and 5, 20, 35, 50, 65, 95, and 125 min after the administration of OA. Each recording session consisted of the simultaneous collection of PAO and V’ coming from 10 or more consecutive breaths in VC-MV, followed by a breath with an inspiratory pause. The pause was held until a stable plateau pressure was reached and was never <2.5 s (Fig. 1).
|
In the reference group of pigs, 2 recording sessions per time interval were performed producing 16 recordings per each of the 16 animals. In the prospective group of pigs, only one recording session per time interval was executed. At the end, we obtained a total of 248 tracings in the reference group of pigs, composing the training pool of curves and a total of 57 recordings from the prospective group of animals (forming the prospective pool of curves).
To introduce a further source of variability during ANN training, during the interval between 65 and 95 min after the administration of OA, PEEPe was randomly changed to 0 or 10 cm H2O. Of the recorded sequence of breaths, the last breath, i.e., the one having the EIP, was used for the manual calculation of CRS by applying the IT (5).
The inspiratory limb of the volume-pressure loop (V-P Loop) of the breath immediately preceding the breath with an inspiratory pause was given to the ANN (Fig. 2).
|
To avoid redundancy of information among neighbor points, the inspiratory limb was under-sampled by taking 50 equally spaced coordinates of PAO and V’. Each curve was re-scaled using as scale factor the value of maximum airway pressure. Each input vector was composed of 100 values (50 x-values plus the corresponding 50 y-values). The ANNs were implemented via software on a computer (MatLab, MathWorks, Natick, MA). The learning algorithm was resilient back propagation. The number of neurons in the input layer was 100; the output layer consisted of 1 neuron, yielding the CRS calculated by the ANN.
To determine the best architecture, i.e., the number of intermediate neurons that provided the best performance for the required task, we used an eight-fold cross-validation with early stopping (more details are given in the on-line data supplement). We tested architectures with a number of intermediate neurons between 2 and 30. The pool of curves composing the reference group of pigs was randomly divided into 2 new subsets in the ratio of 80:20. The biggest subset was used for training the ANN and the smaller one to validate the learning process.
The prospective trial consisted of presenting the tracings coming from the prospective group of 8 pigs (57 recordings) to the chosen ANN. ANN performance was studied by calculating the linear regression between the results yielded by the ANN and the manual measure of CRS (using IT) and analyzing the measurement error according to Bland and Altman (13). The measurement error was calculated over the entire prospective pool of curves and also in healthy and sick conditions of the lung separately. Linear regression was used to analyze whether the error by the ANN was dependent from the absolute level of compliance.
|
Results |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionReferences |
|---|
The first step of the study was to identify the best ANN architecture for the required task; it was the one with 25 intermediate neurons, which showed the least mean squared error (MSE) in the eight cross-validation tests. After obtaining this information, the final training phase could start. One hundred ANNs, all having an identical architecture (100 input, 25 intermediate, and 1 output neurons) but differing for a random assignment of the starting weights to the nodes, underwent the process of training. The one that reached the smallest average MSE during the allowed number of training cycles was chosen for the subsequent experiments (Fig. 3). The best performing ANN showed an average MSE of 0.57 mL/cm H2O; the second best ANN presented an average MSE of 0.62 mL/cm H2O.
|
At the end of the training procedure, the chosen ANN showed a performance expressed by the linear regression y = 0.98x + 0.53 with r = 0.99 (r2 = 0.98), where x is CRS calculated according to IT, and y is CRS extracted by ANN.
In the reference group of pigs, 125 min after the injection of OA, CRS decreased from 22.9 ± 5.5 (at baseline) to 14.8 ± 3.9 mL/cm H2O (Fig. 4). Baseline data and values obtained 5 min after OA injection did not differ and were used as healthy lung (HL) data. CRS was significantly reduced from 20 min after OA injection (P < 0.01), and results from 20 to 125 min after OA injection were used as sick lung (SL) data.
|
The administration of OA in the prospective group of animals produced a decrease in CRS (from 25.0 ± 5.5 mL/cm H2O to 15.8 ± 3.3 mL/cm H2O 125 min later).
Applying the previously trained ANN to the prospective group of data, the ANN performance in assessing CRS was expressed by (using the same notation as above) y = 0.94x + 1.69 with r = 0.97 (r2 = 0.94) (Fig. 5).
|
Bland-Altman analysis of ANN versus IT methods for extracting CRS showed a bias ± SD of -0.67 ± 1.52 mL/cm H2O. Analysis of the ANN performance in HL and SL separately showed a bias ± SD of -0.46 ± 2.14 mL/cm H2O in HL and -0.75 ± 1.26 mL/cm H2O in SL (Fig. 6).
|
Comparing the two populations of errors (error in assessing CRS by ANN in HL and in SL) by the application of the two-tailed Student’s t-test showed no significant difference between the two groups (P = 0.53).
Linear regression of ANN error in assessing CRS (y) versus the average measure of CRS by ANN and IT (x) showed y = 0.03x - 1.23 with r = 0.12 (r2 = 0.01), which was not significant. Thus, the severity in lung pathology, as assessed by CRS, had no influence on the relationship between ANN and IT.
Moreover, the error in estimating CRS from recordings made at PEEPe = 0 and PEEPe = 10 is not different from the error obtained from recordings at PEEPe = 5.
|
Discussion |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionReferences |
|---|
This study showed that CRS can be estimated by ANN in a VC-MV setting without having to stop V’. ANN assessed the CRS with a small error and a small scatter in both HL and SL. The amount of error was not statistically different in HL and SL conditions; the ANN error had no dependency from the absolute level of CRS.
The tracings coming from the animal were presented to the ANN as V-P Loops. The reason for this choice is that the CRS function is the derivative (i.e., the instantaneous slope) of a V-P curve, and it is not influenced by any re-scaling procedure that affects the absolute values of volume and pressure provided that the ratio between them is maintained. We have used this property of derivative functions for applying one of the known heuristic methods to obtain better performances by ANNs. We have separately scaled all curves to make all the pairs of coordinates range between 0 and 1. Each V-P Loop had its maximum of pressure as its own scaling factor. In each loop, all the pairs of coordinates were divided by this scaling factor. This procedure made each curve retain its values of instantaneous compliance but, at the same time, hastened the ANN convergence towards a solution because of the use of smaller numbers (14).
Whereas in the reference group of animals we collected two recordings per time interval, in the prospective group of animals, we collected only one per time interval. This decision depended on the fact that in the training phase, we needed to collect a number of examples as big as possible and also with very slight difference between them. The prospective phase was oriented to eliciting performance error caused by biological variability; it was required to test only one condition per time interval per animal to avoid adding artifacts to the estimation of ANN performance because of multiple similar examples.
The basic advance of this paper over our previous one (12) is the possibility of assessing lung compliance without having to stop V’. This achievement may make the ANN-based method a valuable monitoring tool.
The IT has the advantage of being based on static or semistatic conditions, such as those obtained after flow interruption. After a transient, pressure reaches a plateau that is the expression of the static properties of the respiratory system. During the plateau, the impact of transient phenomena, such as stress-relaxation or gas redistribution, can be considered negligible (15). Thus, to obtain the best results from IT, it is required to maintain the occlusion for some time. The EIP, if it exists at all, may be too short to allow pressure equilibration and the calculation of a true CRS. A deliberate prolongation of the breath hold can raise concerns about the potential for harmful consequences in circulatory unstable patients. Neumann et al. (16) showed, in an animal model of acute lung injury by OA administration, that alveolar recruitment is a continuous event during inspiration, and an EIP may have an impact on the efficiency of such recruitment. Moreover, an end-inspiratory hold maneuver for measuring respiratory mechanics may yield results that are different from the ones obtained during ventilation without EIP.
The necessity of monitoring respiratory mechanics without interfering with the actual pattern of breathing required investigators to apply inverse modeling techniques. One of the most frequently used is the MLF method (1). Application of MLF is based on the assumption that the mechanical properties of the respiratory system are constant and time-invariant over the analyzed breath. This introduces an error in estimating respiratory variables that becomes more important during pathologic conditions. In fact, in these conditions, it is possible to verify that respiratory mechanics may no longer be linear (17), and the accuracy of the method becomes decreased, thus requiring corrections of the algorithm (18) or limiting the application of MLF to specific segments of the breathing cycle (19).
The ANN-based method does not require any assumption regarding the model to be fitted because the type of ANN used in this paper (multilayer perceptrons) are universal function approximators (7,20,21) and can reproduce models of high complexity without limitations to the degree of nonlinearity.
The ANN-based method, as described here, computes only one value of CRS (the one that is referred to as the actual pattern of breathing). It may be disputable whether one value is sufficient for the assessment of the elastic properties of the respiratory system. If the objective is the speculative analysis of the mechanical properties of the lung, the best choice still remains to draw a V-P curve, although it has been affirmed to be difficult in some contingencies (22).
When the objective is to monitor a variable for the control of a machine, one needs a method that is not only accurate, but also robust. Several working definitions of robustness have been proposed; in the case of estimation methods, the adjective "robust" is applied to those methods that "work well not only under ideal conditions, but also under conditions representing a departure from an assumed distribution or model (23)."
MLF can maintain a certain stability in the presence of random noise with Gaussian distribution (24), but in the analysis of multicompartmental models, a variable amount of error will affect the MLF (18). MLF is thus robust only when the respiratory system is in healthy conditions or presents a pathology that uniformly affects the lungs, thus modifying its mechanical properties in a homogeneous way.
ANN-based methods are robust because when extracting information from a curve, they do not require a preconceived model to be fitted. Having no limitation of model equation degree, by definition they are not affected by changes from linear to nonlinear behaviors provided that they have been trained on examples of both situations. This paper does not address the problem of quantification of the robustness of the method; however, there is strong mathematical support in computer science literature on the qualities of ANN in this respect (25).
ANN-based technology, although theoretically applicable to different clinical conditions, in this paper has been tested on a limited number of conditions. In future studies, it will be required to measure ANN performance in other settings, such as different breathing patterns, or other lung pathologies not considered here. The advantage of ANN technology, as emphasized here, is that it does not require explicit programming. The disadvantage is that ANN requires examples to be trained on.
|
Acknowledgments |
|---|
Supported, in part, by the Swedish Medical Research Council (5315); The Swedish Heart-Lung Fund; The School of Anesthesiology and Intensive Care Medicine, Bari University, Italy; The Center of Innovative Technologies for Signal Detection and Processing (TIRES), Bari University, Italy.
The authors wish to thank Karin Fagerbrink, Eva-Maria Hedin, and Agneta Roneus, laboratory assistants at the Department of Clinical Physiology, Uppsala University, for their invaluable contribution to the success of these experiments.
|
References |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionReferences |
|---|
- Iotti G, Braschi A, Brunner J, et al. Respiratory mechanics by least squares fitting in mechanically ventilated patients: applications during pressure support ventilation. Intensive Care Med 1995; 21: 406–13.[Web of Science][Medline]
- Suratt P, Owens D, Kilgore W, et al. A pulse method of measuring respiratory system compliance. J Appl Physiol 1980; 49: 1116–21.
[Abstract/Free Full Text] - Suratt P, Owens D. A pulse method of measuring respiratory system compliance in ventilated patients. Chest 1981; 80: 34–8.
[Abstract/Free Full Text] - Bates JHT, Rossi A, Milic-Emili J. Analysis of the behavior of the respiratory system with constant inspiratory flow. J Appl Physiol 1985; 58: 1840–8.
[Abstract/Free Full Text] - Bates JHT, Baconnier P, Milic-Emili J. A theoretical analysis of interrupter technique for measuring respiratory mechanics. J Appl Physiol 1988; 64: 2204–14.
[Abstract/Free Full Text] - Cross SS, Harrison RF, Kennedy RL. Introduction to neural networks. Lancet 1995; 346: 1075–9.[Web of Science][Medline]
- Hornik K, Stinchcombe M, White H. Multilayer feedforward networks are universal approximators. Neural Networks 1989; 2: 359–66.[Web of Science]
- Orr AJ, Westenskow DR. A breathing circuit alarm system based on neural networks. J Clin Monit 1994; 10: 101–9.[Medline]
- Bright P, Miller MR, Franklyn JA, Sheppard MC. The use of a neural network to detect upper airway obstruction caused by goiter. Am J Respir Crit Care Med 1998; 157: 1885–91.
- Leon MA, Räsänen J, Mangar D. Neural network-based detection of esophageal intubation. Anesth Analg 1994; 78: 548–53.
[Abstract/Free Full Text] - Räsänen J, León M. Detection of lung injury with conventional and neural network-based analysis of continuous data. J Clin Monit 1998; 14: 433–9.
- Perchiazzi G, Högman M, Rylander C, et al. Assessment of respiratory system mechanics by artificial neural networks: an exploratory study. J Appl Physiol 2001; 90: 1817–24.
[Abstract/Free Full Text] - Bland JM, Altman DG. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 1986; 1: 307–10.[Web of Science][Medline]
- Haykin S. Heuristics for making the back-propagation algorithm perform better. In: Haykin S, ed. Neural networks: a comprehensive foundation. London: Prentice-Hall International, 1999: 178–84.
- Similowski T, Levy P, Corbeil C, et al. Viscoelastic behavior of lung and chest wall in dogs determined by flow interruption. J Appl Physiol 1989; 67: 2219–29.
[Abstract/Free Full Text] - Neumann P, Berglund JE, Mondejar EF, et al. Effect of different pressure levels on the dynamics of lung collapse and recruitment in oleic-acid-induced lung injury. Am J Respir Crit Care Med 1998; 158: 1636–43.
[Abstract/Free Full Text] - Ranieri VM, Giuliani R, Fiore T, et al. Volume-pressure curve of the respiratory system predicts effects of PEEP in ARDS: "occlusion" versus "constant flow" technique. Am J Respir Crit Care Med 1994; 149: 19–27.[Abstract]
- Kano S, Lanteri CJ, Duncan AW, Sly PD. Influence of nonlinearities on estimates of respiratory mechanics using multilinear regression analysis. J Appl Physiol 1994; 77: 1185–97.
[Abstract/Free Full Text] - Guttmann J, Eberhard L, Wolff G, et al. Maneuver-free determination of compliance and resistance in ventilated ARDS patients. Chest 1992; 102: 1235–42.
[Abstract/Free Full Text] - Hornik K, Stinchcombe M, White H. Universal approximation of an unknown mapping and its derivatives using multilayer feedforward networks. Neural Networks 1990; 3: 551–60.[Web of Science]
- Haykin S. Approximation of functions: universal approximation theorem. In: Haykin S, ed. Neural networks: a comprehensive foundation. London: Prentice-Hall International, 1999: 208–9.
- Pelosi P, Gattinoni L. Respiratory mechanics in ARDS: a siren for physician? Intensive Care Med 2000; 26: 653–6.[Web of Science][Medline]
- Santa Fe Institute. Working definitions of robustness. Available at: http://discuss.santafe.edu/robustness/stories/storyReader$9. Accessed July 1, 2003.
- Stegmaier PA, Zollinger A, Brunner J, Pasch T. Assessment of pulmonary mechanics in mechanical ventilation: effects of imprecise breath detection, phase shift and noise. J Clin Monit Comput 1998; 14: 127–34.[Medline]
- Kerlirzin P, Vallet F. Robustness in multilayer perceptrons. Neural Comput 1993; 5: 473–82.
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
