| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
1 Department of Emergency and Transplantation, Bari University Hospital, 70124 Bari, Italy; 2 Department of Clinical Physiology, Uppsala University Hospital, S-75185 Uppsala, Sweden; and 3 Department of Anaesthesia, Sahlgrenska University Hospital, S-41345 Göteborg, Sweden
 |
ABSTRACT |
|---|
 TOP
 ABSTRACT
 INTRODUCTION
MATERIALS AND METHODS
RESULTS
DISCUSSION
REFERENCES
|
|---|
We evaluated 1) the performance of an artificial neural network (ANN)-based technology in assessing the respiratory system resistance (Rrs) and compliance (Crs) in a porcine model of acute lung injury and 2) the possibility of using, for ANN training, signals coming from an electrical analog (EA) of the lung. Two differently experienced ANNs were compared. One ANN (ANNBIO) was trained on tracings recorded at different time points after the administration of oleic acid in 10 anesthetized and paralyzed pigs during constant-flow mechanical ventilation. A second ANN (ANNMOD) was trained on EA simulations. Both ANNs were evaluated prospectively on data coming from four different pigs. Linear regression between ANN output and manually computed mechanics showed a regression coefficient (R) of 0.98 for both ANNs in assessing Crs. On Rrs, ANNBIO showed a performance expressed by R = 0.40 and ANNMOD by R = 0.61. These results suggest that ANNs can learn to assess the respiratory system mechanics during mechanical ventilation but that the assessment of resistance and compliance by ANNs may require different approaches.
resistance; compliance; oleic acid; acute lung injury
| |
INTRODUCTION |
|---|
|
TOP
ABSTRACT
INTRODUCTION
MATERIALS AND METHODS
RESULTS
DISCUSSION
REFERENCES
|
|---|
ARTIFICIAL NEURAL NETWORKS (ANNs) are claimed to be systems that enable the execution of a particular task without the need for an a priori knowledge of it. For example, an ANN is able to reproduce the mechanics of a system after having faced a certain number of examples. In recent years, an increasing amount of literature has been published on these methods (5, 9, 10). In the medical field, they can be used in two different ways: as computer-aided decision tools (3, 8) and in signal analysis (23). A large variety of tracings has been studied using ANNs: electrocardiograms (14), electromyograms (1), electroencephalograms (16), arterial pulse waveforms (2), and evoked potentials in anesthesia (15); a review of the published literature is at high risk of incompleteness. However, in the area of respiratory mechanics, little work has been done.
Leon and Lorini (18) investigated the capability of ANNs to identify spontaneous and pressure-support ventilation modes from gas flow and airway pressure signals. Wilks and English (29) used ANNs, in an exploratory experiment, to classify the efficiency of respiratory patterns to predict harmful changes of the O2 saturation in infants. Snowden et al. (27) fed an ANN with blood-gas parameters and the ventilator settings that determined them to obtain advice for new ventilator settings. Orr, Westenskow, and colleagues (21, 28) studied the use of intelligent alarms based on ANNs for anesthesia breathing circuits. Bright et al. (7) described the use of an ANN to identify upper airway obstruction. The ANN was fed with six indexes taken from the expiratory limb of a flow-volume loop, and the performance obtained was better than human experts at identifying flow loops with upper airway obstruction. Leon et al. (19) developed a successful ANN-based system to detect esophageal intubation using airway flow and pressure signals. Räsänen and León (25), with the respiratory tracings of healthy and oleic acid-injured lungs of dogs, trained an ANN to assess the presence and extent of lung damage.
The aims of the experiments reported here were to evaluate in an animal
model 1) whether ANNs can assess respiratory system resistance (Rrs) and compliance (Crs) using the tracings of airway opening pressure (Pao), instantaneous inspiratory flow
(
I) and tidal volume (VT), during an
end-inspiratory hold maneuver (e-IHM), and 2) whether it is
possible to substitute animal tracings, in the learning process, with
simulations obtained by nonbiological models.
| |
MATERIALS AND METHODS |
|---|
|
TOP
ABSTRACT
INTRODUCTION
MATERIALS AND METHODS
RESULTS
DISCUSSION
REFERENCES
|
|---|
Experimental design.
The study was divided into three parts. The first phase intended to
train an ANN to extract Rrs and Crs from the tracing of Pao,
VT, and I, using data from an animal
model of acute lung injury (ANNBIO). The progressive
worsening of the lung after the induction of damage yields different
"snapshots" of the respiratory mechanics as reflected by Rrs and
Crs changing over the time (see Fig. 1).
These snapshots were the examples necessary to train the ANN. Ten pigs
were used to provide this pool of data (the reference group). An expert
made a manual calculation of Rrs and Crs directly on all curves. This
was then used as a reference to which the performance of the ANN was
compared.
|
; rs = Rrs · Crs) as found
in the reference group, different sessions of mechanical ventilation
were simulated. Then, the Pao, VT, and I
collected from the model were used to feed an ANN (ANNMOD)
having the same architecture as the one applied to animal tracings. The
performances of the two networks in assessing the Crs and Rrs were
separately assessed.
In the third phase, the two previously trained ANNs
(ANNBIO and ANNMOD) had to face
prospectively the traces coming from four pigs (the prospective group).
In this way, the performance of the two differently trained ANNs was
compared in assessing the mechanics of new individuals.
The animal model.
Fourteen healthy pigs (Yorkshire, Hampshire, and Swedish mixed country
breed; weight 30.6 ± 6.0 kg) were, after approval by the local
animal ethics committee, included in the study: ten animals to build
the database of respiratory tracings and four animals to make up the
prospective trial group. After premedication with azaperone (40 mg im,
Stresnil, Janssen, Vienna, Austria), given at the farm before
transportation, the pigs underwent general anesthesia, induced by using
atropine (0.04 mg/kg), tiletamin-zolazepam (5 mg/kg Zoletil, Boehringer
Ingelheim, Copenhagen, Denmark), and medetomidin (5 µg/kg, Dormitor
Vet., Orion Pharma, Sollentuna, Sweden), injected intramuscularly in
the neck. Next, they were orally intubated using a cuffed endotracheal
tube (no. 6, Hi-Contour, Mallinckrodt Medical, Athlone, Ireland).
Anesthesia was maintained by a continuous intravenous infusion of
ketamine (20 mg · kg
1 · h1,
Ketaminol, Vetpharma, Zurich, Switzerland), pancuronium (0.24 mg · kg1 · h1, Pavulon,
Organon Teknika, Göteborg, Sweden), and fentanyl (5 µg · kg1 · h1,
Pharmalink, Spånga, Sweden). A constant infusion of Rehydrex (Fresenius Kabi, Uppsala, Sweden) with glucose 25 mg/ml was given at
the average rate of 5 mg · kg1 · h1. Fluid
replacement was constantly regulated to maintain a constant hemoglobin
value and a stable systemic arterial pressure. For invasive arterial
pressure measurement and arterial blood- gas analysis, an 18-gauge
catheter was inserted into the left carotid artery. A floating-tip
pulmonary artery (PA) catheter, together with an 18-gauge catheter, was
introduced into the right external jugular vein. The position of the PA
catheter was confirmed by the pressure traces on the connected bedside
monitor (CS/3, Datex Ohmeda, Helsinki, Finland). Measurement of
oxyhemoglobin saturation (SaO2) was performed by a
transcutaneous sensor (EarSat, Datex Ohmeda). Cardiac output was
measured in triplicate by injection of a cold bolus (~10 ml,
3-5°C) of physiological saline randomly during the respiratory
cycle. A temperature sensor located in the PA catheter allowed
continuous monitoring of blood temperature.
Ventilation. The pigs were ventilated using volume-controlled constant-flow mechanical ventilation (Servo 900 C, Siemens Elema, Solna, Sweden). Inspiratory-to-expiratory ratio was set to 1:2 (s); the initial VT was set at 15 ml/kg, extrinsic positive end-expiratory pressure (PEEPe) was initially set to 5 cmH2O. Blood gases were used to adjust the VT to result in normocapnia (35-45 Torr). Inspiratory fraction of oxygen was kept at 0.5, with the exception of a predefined set of five breaths at an inspiratory oxygen fraction of 1.0 before inspiratory hold maneuvers if SaO2 was <80%.
Lung injury. Oleic acid (OA) 0.1 ml/kg (Apoteksbolaget, Göteborg, Sweden) was fractionally injected directly into the central venous catheter in repeated doses of 0.5 ml. Before entering the circulation, it was mixed in a three-way stopcock with the turbulent flow of a high-pressure washing line opened during the injection. This procedure allowed as complete as possible dispersion of the OA into the infusate, avoiding large droplets. Administration of OA was suspended if SaO2 fell to 80%. Any fall in systemic arterial pressure during OA injection was countered using epinephrine, in boluses of 0.01 mg.
Respiratory variables recording. A D-Lite connector (Datex Ohmeda) was mounted to the endotracheal tube. The two sampling ports of the D-Lite were connected to a differential pressure transducer (Sensym, SensorTechnics, Pucheim, Germany). The transducer was calibrated at the beginning of each experimental session, using a water column for static pressures. Calibration of flow measurement was performed with a source of constant flow, the transducer to be calibrated and a precision flowmeter (calibration analyzer TS4121/P, Timeter Instrument, St. Louis, MO) connected in series. Moreover, by using a pneumatic short-circuit system (20) driven by a magnetic valve, it was possible to check the transducer for the zero point value at the beginning and repeatedly during the experimental sessions. Data from the transducer were collected by the Carina 2.4.0 acquisition program (C-O Sjöberg Engineering, Upplands-Väsby, Sweden), purposely written for the LabView acquisition system (LabView 4.0.1, National Instruments, Austin, TX). The traces of flow and pressure, collected in real time at 200 Hz, were stored on the hard disk of a personal computer. Pressure and flow tracings were recorded after a stabilization period of 60 min following instrumentation and 5, 20, 35, 50, 65, 95, and 125 min after the first administration of OA. Each recording comprised duplicate e-IHMs separated by 10 or more normal breaths. By this means, 16 curves per animal were recorded. In the period between 65 and 95 min after the first administration of OA, the PEEPe of 5 cmH2O was randomly changed to 0 or 10 cmH2O. After that, PEEPe was returned to 5 cmH2O. This intervention was introduced to provide a further source of variability during the processes of training and testing the ANNs.
Measurement of Rrs and Crs.
From the recorded curves, it was necessary to obtain Rrs and Crs
independently from the ANNs to have a standard of comparison. Crs was
calculated as the ratio between VT and 
pst,
where pst = P2 
PEEPe, where
P2 is the pressure recorded after 2 s of
end-inspiratory hold (4). Calculation of VT
was performed by integration of the inspiratory flow. Rrs was defined
by the drop in pressure (pdyn) divided by
I, where pdyn = Ppeak P1 (4).
Ppeak is the maximum pressure that is reached just before
the e-IHM and P1 is the level of pressure as soon as flow
is stopped by the e-IHM. The measurements were corrected for the
closing time of the ventilator valves according to Kochi et al.
(17). The correct choices of P1 and
I were facilitated by the fact that Pao and
inspiratory flow traces were recorded synchronously with a common time
code. No corrections were done for temperature, humidity, and intrinsic
PEEP. Rrs calculated in this way also includes the tracheal tube resistance.
Preprocessing of raw data.
The recording of Pao during e-IHM, from the PEEP preceding the breath
to 2 s after the e-IHM (comprising 3.2 s of recording), was
submitted to the ANN. The traces were resampled at 50 Hz to avoid
redundancy of information among neighbor points (passing from 640 to
160 points per curve). Each pattern to be analyzed by the ANN comprised
the 160 points described, VT, and I,
arriving to a final input vector of 162 points (see Fig.
2). Both input and output vectors were
rescaled to obtain values that ranged between 0 and 1. When the ANN
yielded its computation of Rrs and Crs, it was remultiplied by the
scale factor to get the measure in centimeters of H2O times
seconds per liter and milliliters per centimeters of H2O,
respectively.
|
Neural networks.
A series of feed-forward ANNs were implemented via software on a
computer (Neural Networks Toolbox ver. 3.0 for MatLab ver. 5, The
MathWorks, Natick, MA). The learning algorithm was resilient backpropagation, using as increment change (
) a value of 1.2 and as
decrement a value of 0.5. They consisted of three layers, whose
activating functions were log-sigmoids for the input and intermediate
layer and linear for the output layer. The numbers of neurons in the
input and output layers were conditioned by the dimensions of the input
pattern and by the number of expected variables from the net. In this
way, the input layer was composed of 162 neurons and the output layer
of 2 neurons, one yielding Rrs and the other one Crs.
Choice of the best architecture. To choose the best architecture of the ANN, i.e., the number of intermediate neurons that provided the best performance for the required task (see below), a series of tests was performed. These included training several ANNs differing from each other by the number of intermediate neurons. Architectures having all the possible finite numbers of intermediate neurons between 2 and 30 were studied, reaching a total of 29 different architectures. The technique used was a multifold cross-validation method with early stopping (for details, see Haykin, Ref. 11). The pool of curves obtained by the reference group was composed of 160 traces in random order and was divided into eight groups. Each ANN was trained on seven groups and validated on the last one. This procedure was done eight times for each net, leaving out a different group for the validation. Each training process had to stop if the error on the validation set rose. The performance of each ANN was computed eight times, once per training/validation session. It was assessed by calculating the mean squared error (MSE) of the ANN on Rrs and Crs. Moreover, it was possible to use the average MSE of the best network as our target MSE for ANNBIO final training.
Training of ANNBIO. After the choice of architecture and target MSE to aim at, it was necessary to train the ANN to assess respiratory system mechanics. The pool of data coming from the reference group (160 curves) was randomly divided in two new subsets. One made up the final training pool composed of 128 curves (80% of the reference group data), and the other formed the final test pool (TeBIO-P) with 32 curves (20% of reference group data). The training stopped if the performance goal (lowest MSE, produced by the best architecture, in the previous experimental phase) was met, if the number of presentations of the training set exceeded 1,000 iterations, or if the MSE on TeBIO-P started to rise (early stopping). Fifty different ANNs, having the same architecture but different node weights, were trained. The one having the lowest MSE on TeBIO-P was eligible for the prospective test. Assessment of training was performed by measuring the final MSE on TeBIO-P and by computing linear regression between results from ANN output and those manually calculated.
Electrical analog of the lung.
An electrical analog of the lung was implemented on a computer via
software (Pater program, University of Bari, Italy). It reproduced the
model proposed by Otis et al. (22) in 1956 and was
composed of two parallel limbs connected to the equivalent of airway
opening by a common resistance. Each limb consisted of a resistor and a
capacitor, connected in series. This circuit is a bicompartmental model
of the lung. By changing the values of resistance and compliance,
different conditions were simulated. "Ventilation" of the model is
obtained by simulating the application of forcing currents, whose
variation in time is equal to the ventilatory pattern. Resistance and
compliance were kept constant and invariable over the time during each
ventilatory session. Each session consisted of six respiratory cycles
to ensure that the last cycle was performed under steady-state
condition. The values assigned to each resistance and compliance were
chosen to reproduce the same range of Rrs, Crs, and time constant
(rs = Rrs · Crs) as found in the reference group of
pigs (see Fig. 3). The conditions
simulated were as follows: Rrs varying from 8 to 48 cmH2O · s · l1 in steps of
five and Crs from 9 to 49 ml/cmH2O in steps of five. Inspiration lasted 1 s, followed by an inspiratory pause of 2 s and an expiration of 2 s. Delivered flow varied from 0.25 to 0.5 l/s in steps of 0.05. PEEP was set to 0, 5, and 10 cmH2O. The combinations of Rrs and Crs whose was either <100 ms or >900
ms were discarded. Combining the mechanical parameters and the
ventilatory patterns, we obtained 750 simulations. From each session,
the recordings of the last cycle were taken and fed to the ANNs. At the
end of the procedure we had 750 tracings that made up the electrical
analog pool of data (EA-P).
|
Training of ANNMOD. ANNMOD had the same architecture as ANNBIO. To choose the ones with the best performances, fifty ANNs were trained on the EA-P. They were trained using the same stopping conditions as the ones imposed on ANNBIO. EA-P was randomly divided into two subsets: a training pool composed of 600 curves (80% of EA-P) and a test pool (TeMOD-P) of 150 curves (20% of EA-P). Its training was quantified on the performance that ANNBIO showed on TeMOD-P by measuring MSE and by computing linear regression between its output and the manually computed results. The ANN having the lowest MSE was eligible for the prospective test.
Prospective test. After the two ANNs (ANNBIO and ANNMOD) were trained and tested on the two different data sets, they had to face the prospective test, which consisted of feeding them the tracings from the four pigs in the prospective group. The prospective pool of curves was composed of 64 tracings (16 per pig). The performance of the two ANNs was studied by calculating MSE (reported in scaled units) and by linear regression between ANN output and manually computed data. Moreover, we calculated the error of the measurements (difference, measure by measure between ANN-yielded data and manually calculated values), computing its mean (bias of the measure) and SD. A Bland-Altman plot for agreement between measurements using these data was drawn (6).
| |
RESULTS |
|---|
|
TOP
ABSTRACT
INTRODUCTION
MATERIALS AND METHODS
RESULTS
DISCUSSION
REFERENCES
|
|---|
Manually computed mechanics.
Crs dropped from 28.5 ± 7.4 to 16.5 ± 4.3 ml/cmH2O and Rrs rose from 13.8 ± 2.8 to 19.7 ± 8.8 cmH2O · s · l 1 over a
125-min period after first OA injection (means ± SD). In
particular, pigs belonging to reference group showed Crs decreasing from 30.2 ± 6.4 to 16.9 ± 3.7 ml/cmH2O and Rrs
increasing from 14.4 ± 3.0 to 23.1 ± 9.0 cmH2O · s · l 1. Pigs in the
prospective group had a Crs that changed from 24.0 ± 8.2 to
15.6 ± 5.5 ml/cmH2O and a Rrs from 12.1 ± 1.2 to 12.9 ± 1.5 cmH2O · s · l1 (see Fig.
4).
|
Choice of ANN architecture. The architecture having the lowest average MSE over the eight sessions of cross-validation test was the one with 27 intermediate neurons. It reached an average MSE of 0.0006.
ANN training.
After training was completed, the performance of the ANNBIO
on TeBIO-P was expressed by a MSE of 0.00081. The linear
regression between neural network output and manually computed values
had a regression coefficient (R) of 0.97 for Rrs; for Crs it
was R = 0.99. ANNMOD, after training,
showed a MSE of 0.00056 on TeMOD-P. Linear regression
between ANNMOD output and manually calculated values had R = 0.99 on Rrs and R = 1.00 on Crs (see Table 1).
|
Prospective test. Linear regression between measured respiratory mechanics by ANNBIO and manually computed values was expressed by R = 0.40 on Rrs and R = 0.98 on Crs with MSE = 0.0037. For ANNMOD, regression was expressed by R = 0.61 on Rrs and R = 0.98 on Crs with MSE = 0.0038.
The measurement errors, expressed as means ± SD, were2.11 ± 4.61 cmH2O · s · l1 on Rrs and
0.17 ± 1.14 ml/cmH2O on Crs for
ANNBIO. Values for ANNMOD were 0.36 ± 4.81 cmH2O · s · l1 on Rrs
and 1.57 ± 1.31 ml/cmH2O on Crs. (See Fig.
5 and Table 1.)
|
| |
DISCUSSION |
|---|
|
TOP
ABSTRACT
INTRODUCTION
MATERIALS AND METHODS
RESULTS
DISCUSSION
REFERENCES
|
|---|
The best ANN architecture for the assessment of respiratory system mechanics was chosen by first using a pool of data (from the reference group of pigs). Then, the ANNs' training started. Fifty ANNs were trained and tested using biological data from the reference group, and another fifty were trained and tested on data from the electrical analog of the lung. All had the same architecture (i.e., the same number and the same disposition of nodes). The ANN with the lowest error on biological tracings and the ANN with the lowest error on electrical analog tracings were picked out. Finally, these two ANN faced data coming from different individuals (the prospective group), and their performance in assessing lung mechanics was separately assessed. In both ANNs, the performance on the prospective group was very good for the assessment of Crs and not as good for the assessment of Rrs.
About the methods. The decision to avoid corrections for temperature, humidity, intrinsic PEEP, and endotracheal tube resistance is a consequence of the aims of this study. We wanted to evaluate whether the ANNs "saw what we could see" in the traces. All the information that was not directly transferable to the ANN was not used in the computation.
Scaling the input and the output vectors of an ANN is a transparent procedure. In fact, the values that the ANNs yielded were multiplied again for the scale factor. Scaling all the input vectors to obtain "small numbers" (namely, having mean values of the input close to zero) is claimed to hasten the learning process (12). We introduced an electrical analog of the lung, mathematically implemented on a computer, to evaluate its usefulness in training an ANN. It is a noise-free source of data that can simulate infinite combinations of Rrs and Crs. By this means, it is possible to obtain tracings created by combinations of Rrs and Crs that, although possible, are unlikely to be found in the spectrum of natural variability. In studies of these new methods, models have great importance. ANNs learn by experience. However, in the field of experimental biology, it is not so easy to obtain the large number of examples that are required to train ANNs. Models can provide the number of examples necessary for training both in the normal range of variability and in extreme conditions. The process of choosing the best "physical" structure of an ANN (i.e., the architecture) is different from the process of training it. After we chose the architecture of the ANN, the only information we used in the successive phases of the experiments was the number and disposition of nodes plus the MSE that was reached. This value of MSE was used as the performance goal at which we could aim during the effective training phase. Dimensioning the architecture of the ANNs was accomplished on biological tracings. The reason for this choice lies in the fact that any ANN has to be dimensioned on the most complicated task that is foreseeable during its use. In fact, real recordings contain different forms of biological noise, which makes the task of extracting lung mechanics variables more complex compared with the same task performed on simulated tracings. After we decided their architecture in the previous part of the study, 50 ANNs were trained on biological data and 50 on model tracings. Each ANN started with random weights assigned to its connections, and the process of learning consisted of adjusting them. Although two networks having the same architecture can show similar performance on the same task, in reality they start the process of learning from different "backgrounds." This background is the random assignment of weights to the ANN when it is initialized. This means that, to observe similar performance by two physically similar ANNs, it is necessary to push the number of iterations to a high number. To optimize our computational resources, we preferred to stop the training after a discrete number of iterations and evaluate the performance. The ANNs with the lower MSE were the ones in which the convergence to a minimum error was also driven by a favorable random matrix of weights assigned at the beginning.About the results. ANNBIO and ANNMOD were able to learn the relation between the input pattern and the corresponding Rrs and Crs. This fact is demonstrated by the performance showed on their respective test groups (TeMOD-P and TeBIO-P), expressed by R of no lower than 0.97. However, we did not consider these results sufficient to get any conclusion about their possible use. Hence, the decision for a prospective study was made. As a result, in these prospective tests, the performance on Crs remained very good, both by ANNBIO and ANNMOD. However, this was not true for Rrs. In fact, the regression coefficient of ANN performance vs. manually computed results was 0.98 for Crs (in both networks) but only 0.40 (in ANNBIO) and 0.61 (in ANNMOD) for Rrs.
The values of Rrs and Crs expressed by the reference and the prospective group of animals may belong to different statistical populations (Fig. 3). However, the problem of statistical difference between the groups has not been addressed in this paper because it was not required by the study. ANNs perform better when they face data that are internal to the range of variation of the data used for training. Whether the two groups are statistically different has no importance. In fact, each tracing of the prospective group is a single entity that is given alone to the network and to which the network answers with an output based on its previously acquired knowledge. This means that the sequence of curves or their characteristics do not affect the performance of an ANN on a successive one. As can be seen in Fig. 3, the measures of Rrs and Crs in the prospective group are reported to be near the lower limits of the variability spectrum. Moreover, ANNBIO was also tested on a few values (equal to 3% of the 64 measurement composing the prospective group) that were outside the range of its training. But this fact cannot explain the poor regression between Rrs measurements. More probably, the reason for poorer performance on Rrs lies in the architecture of the net. In fact, the strength of ANNs and, in general, of processing systems distributed in parallel is attributed to the capability of extracting information from numerous data. Thus, if a part of these vectors differs from the prototype used for training, the overall performance should not be much affected. To extract Crs, ANNs can use all the points from P1 to P2 (2 s, for a total of 100 points at 50 Hz sampling). We can consider also information present in the slope of Pao during inspiration (~1 s, another 50 points) because flow is constant and the ratio of volume to Pao, although in dynamic conditions, is a measure of Crs. Information regarding resistance is only in two small groups of points. These are the drop between Ppeak and P1 and also the very first rise in pressure, at the beginning of the flow (the so-called resistive rise). It comprises only 20-30 points. If the signal presents any source of error, because each point contributes only for a fraction to the final calculation, calculations based on more points have bigger probabilities of compensating for an error. Considering that resistance in this particular setting is calculated only on a few points (in comparison to compliance), the possibilities of being affected by whatever kind of noise are greater than for compliance. For future development of ANN in the evaluation of respiratory mechanics, the use of two different ANNs may be considered: one for Crs and another for Rrs. In fact, we must consider that a global error function such as MSE computes the error on both Rrs and Crs. And so, for a given target MSE, if Crs is perfectly assessed by the ANN, the results on Rrs could also be less good, on condition that overall MSE is within the imposed limit. This consideration brings another important point into the discussion. It is the idea that different ANN architectures (and indirectly the way they are fed by information) could have different performance on the same task. We wanted to explore the practical problems of their use in our field of research. To perform this particular task, it was necessary to reduce to the minimum the preprocessing of curves (hence the decision of giving to the ANNs the complete Pao curve) and to train the ANNs to calculate Rrs and Crs at the same time and in the same way. Moreover, we reduced all the operations requiring choices by human "experts" (i.e., cutting parts of the trace, or picking only significant points). They could have shown better performances if, instead of complete traces, we gave only some characteristics of the curves. However, in this case, we would not have had the possibility of evaluating the robustness of ANNs to noise and perturbation of signals. In fact, adding the step of human choice in picking what is important and what is not, we would have also added another source of error. Another important message arrives from the comparison between ANNBIO and ANNMOD. Electrical analogs of the lung can be used in the training phase of ANNs. The two differently experienced ANNs have largely similar performances in the assessment of Crs. This suggests that training the ANNs can also be done by the use of models, hence the possibility of cutting the time required for training them.Perspectives.
The most important achievement of this study is that an ANN able to
extract Crs from Pao and I signals has been built
up. Now the question is whether we should use these methods or whether there is any case for using ANNs instead of usual techniques. It must
be clear that, when we have the equation to solve a problem and good
data to assign to equation parameters, ANNs are not the best
choice. Their limit is that they may have less precision (although well trained) if they compete against a defined equation and
pure data. So, for research purposes, manual calculation of respiratory
variables may still remain as the reference method. This means that we
may not have the need of using ANNs as an "instrument of
comprehension" in research. In fact, a trained ANN is a black box:
able to reproduce the mechanics of a system but unable to give plain
explanation in what way it does it. This is also the key point when
comparing ANNs to methods of function fitting. In fact, curve fitting
is an analytical method, requiring a preconceived idea about the number
of variables and the complexity of the equation to be applied for
solving a particular problem.
| |
ACKNOWLEDGEMENTS |
|---|
We thank Karin Fagerbrink, Eva-Maria Hedin, and Agneta Roneus, laboratory assistants at Dept. of Clinical Physiology, Uppsala University, for invaluable contribution to the success of these experiments.
| |
FOOTNOTES |
|---|
This study was supported by grants from School of Anaesthesiology and Intensive Care Medicine, Bari University, Italy; The Swedish Medical Research Council (5315); and The Swedish Heart-Lung Fund.
Address for reprint requests and other correspondence: G. Perchiazzi, Istituto di Anestesia e Rianimazione, Universita' di Bari, Ospedale Policlinico, Piazza Giulio Cesare, 70124 Bari, Italy (E-mail: gperchiazzi{at}yahoo.com).
The costs of publication of this article were defrayed in part by the payment of page charges. The article must therefore be hereby marked "advertisement" in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.
Received 15 March 2000; accepted in final form 21 December 2000.
| |
REFERENCES |
|---|
|
TOP
ABSTRACT
INTRODUCTION
MATERIALS AND METHODS
RESULTS
DISCUSSION
REFERENCES
|
|---|
1.
Abel, EW,
Zacharia PC,
Forster A,
and
Farrow TL.
Neural network analysis of the EMG interference pattern.
Med Eng Phys
18:
12-17,
1996[Medline].
2.
Allen, J,
and
Murray A.
Comparison of three arterial pulse waveform classification techniques.
J Med Eng Technol
20:
109-114,
1996[Medline].
3.
Ashutosh, K,
Lee H,
Mohan CK,
Ranka B,
Mehrotra K,
and
Alexander C.
Prediction criteria for successful weaning from respiratory support: statistical and connectionist analyses.
Crit Care Med
20:
1295-1301,
1992[ISI][Medline].
4.
Bates, JHT,
Rossi A,
and
Milic-Emili J.
Analysis of the behavior of the respiratory system with constant inspiratory flow.
J Appl Physiol
58:
1840-1848,
1985
5.
Baxt, WG.
Application of artificial neural networks to clinical medicine.
Lancet
346:
1135-1138,
1995[ISI][Medline].
6.
Bland, JM,
and
Altman DG.
Statistical methods for assessing agreement between two methods of clinical measurement.
Lancet
1:
307-310,
1986[ISI][Medline].
7.
Bright, P,
Miller MR,
Franklyn JA,
and
Sheppard MC.
The use of a neural network to detect upper airway obstruction caused by goiter.
Am J Respir Crit Care Med
157:
1885-1891,
1998
8.
Buchman, TG,
Kubos KL,
Seidler AJ,
and
Siegfort MJ.
A comparison of statistical and connectionist models for the prediction of chronicity in a surgical intensive care unit.
Crit Care Med
22:
750-762,
1994[ISI][Medline].
9.
Cross, SS,
Harrison RF,
and
Kennedy RL.
Introduction to neural networks.
Lancet
346:
1075-1079,
1995[ISI][Medline].
10.
Dybowski, R,
and
Gant V.
Artificial neural networks in pathology and medical laboratories.
Lancet
346:
1203-1207,
1995[ISI][Medline].
11.
Haykin, S.
Cross-validation.
In: Neural Networks: A Comprehensive Foundation London: Prentice-Hall International, 1999, p. 213-218.
12.
Haykin, S.
Heuristics for making the back-propagation algorithm perform better.
In: Neural Networks: A Comprehensive Foundation. London: Prentice-Hall International, 1999, p. 178-184.
14.
Hedén, B,
Ölin H,
Rittner R,
and
Edenbrandt L.
Acute myocardial infarction detected in the 12-lead ECG by artificial neural networks.
Circulation
96:
1798-1802,
1997
15.
Huang, J, H,
Lu Y,
Nayak A,
and
Roy R J.
Depth of anesthesia estimation and control.
IEEE Trans Biomed Eng
46:
76-81,
1999.
16.
Jandó, G,
Siegel RM,
Horváth Z,
and
Buzsáki G.
Pattern recognition of the electroencephalogram by artificial neural networks.
Electroencephalogr Clin Neurophysiol
86:
100-109,
1993[Medline].
17.
Kochi, T,
Okubo S,
Zin WA,
and
Milic-Emili J.
Flow and volume dependence of pulmonary mechanics in anesthetized cats.
J Appl Physiol
64:
441-450,
1988
18.
Leon, MA,
and
Lorini FL.
Ventilation mode recognition using artificial neural networks.
Comput Biomed Res
30:
373-378,
1997[Medline].
19.
Leon, MA,
Räsänen J,
and
Mangar D.
Neural network-based detection of esophageal intubation.
Anesth Analg
78:
548-553,
1994
20.
Meriläinen, P,
Hänninen H,
and
Tuomaala L.
A novel sensor for routine continuous spirometry of intubated patients.
J Clin Monit
9:
374-380,
1993[ISI][Medline].
21.
Orr, AJ,
and
Westenskow DR.
A breathing circuit alarm system based on neural networks.
J Clin Monit
10:
101-109,
1994[Medline].
22.
Otis, AB,
Mckerrow CB,
Bartlett RA,
Mead J,
Mcilroy MB,
Selverstone NJ,
and
Radford EP.
Mechanical factors in distribution of pulmonary ventilation.
J Appl Physiol
8:
427-443,
1956
23.
Patil, S,
Henry JW,
Rubenfire M,
and
Stein PD.
Neural network in the clinical diagnosis of acute pulmonary embolism.
Chest
104:
1685-1689,
1993
24.
Ranieri, VM.
Optimization of patient-ventilator interactions: closed-loop technology to turn the century.
Intensive Care Med
23:
936-939,
1997[Medline].
25.
Räsänen, J,
and
León M.
Detection of lung injury with conventional and neural network-based analysis of continuous data.
J Clin Monit
14:
433-439,
1998.
27.
Snowden, S,
Brownlee KG,
Smye SW,
and
Dear PRF
An advisory system for artificial ventilation of the newborn utilizing a neural network.
Med Inform (Lond)
18:
367-376,
1993[Medline].
28.
Westenskow, DR,
Orr AJ,
Simon FH,
Bender HJ,
and
Frankenberger H.
Intelligent alarms reduce anesthesiologist's response time to critical faults.
Anesthesiology
77:
1074-1079,
1992[Medline].
29.
Wilks, PAD,
and
English MJ.
A system for rapid identification of respiratory abnormalities using a neural network.
Med Eng Phys
17:
551-555,
1995[Medline].
This article has been cited by other articles:
|
|
 |
|
  S. H. Kim, J. M. Lee, J. H. Kim, K. G. Kim, J. K. Han, K. H. Lee, S. H. Park, N.-J. Yi, K.-S. Suh, S. K. An, et al. Appropriateness of a Donor Liver with Respect to Macrosteatosis: Application of Artificial Neural Networks to US Images--Initial Experience Radiology, March 1, 2005; 234(3): 793 - 803. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 M. Nirmalan, M. Niranjan, T. Willard, J. D. Edwards, R. A. Little, and P. M. Dark Estimation of errors in determining intrathoracic blood volume using thermal dilution in pigs with acute lung injury and haemorrhage Br. J. Anaesth., October 1, 2004; 93(4): 546 - 551. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 G. Perchiazzi, R. Giuliani, L. Ruggiero, T. Fiore, and G. Hedenstierna Estimating Respiratory System Compliance During Mechanical Ventilation Using Artificial Neural Networks Anesth. Analg., October 1, 2003; 97(4): 1143 - 1148. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| Visit Other APS Journals Online |
,
Respiratory mechanics of a reference group of pigs; 
,
respiratory mechanics of a prospective group of pigs; ×, implemented
simulations. Cmin = 9 ml/cmH2O;
Cmax = 49 ml/cmH2O; Rmin = 8 cmH2O · s · l
