- Email: [email protected]
Intelligent model-based advisory system for the management of ventilated intensive care patients. Part II: Advisory system design and evaluation
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 9 9 ( 2 0 1 0 ) 208–217
journal homepage: www.intl.elsevierhealth.com/journals/cmpb
Intelligent model-based advisory system for the management of ventilated intensive care patients. Part II: Advisory system design and evaluation Ang Wang a , Mahdi Mahfouf b,∗ , Gary H. Mills c , G. Panoutsos b , D.A. Linkens b , K. Goode d , Hoi-Fei Kwok e , Mouloud Denaï a a
Process Automation, ABB Limited, Howard Road, Eaton Socon, Cambridgeshire PE19 8EU, UK Department of Automatic and Systems Engineering, The University of Sheffield, Mappin Street, Sheffield S1 3JD, UK c Department of Critical Care and Anaesthesia, Northern General Hospital, Herries Road, Sheffield S5 7AU, UK d Postgraduate Medical Institute, The University of Hull, Cottingham Road, Hull HU6 7RX, UK e School of Psychology, The University of Birmingham, Edgbaston, Birmingham B15 2TT, UK b
a r t i c l e
i n f o
a b s t r a c t
Article history:
The optimisation of ventilatory support is a crucial issue for the management of respira-
Received 10 April 2009
tory failure in critically ill patients, aiming at improving gas exchange while preventing
Received in revised form
ventilator-induced dysfunction of the respiratory system. Clinicians often rely on their
15 March 2010
knowledge/experience and regular observation of the patient’s response for adjusting the
Accepted 15 March 2010
level of respiratory support. Using a similar data-driven decision-making methodology, an adaptive model-based advisory system has been designed for the clinical monitoring
Keywords:
and management of mechanically ventilated patients. The hybrid blood gas patient model
Mechanical ventilation
SOPAVent developed in Part I of this paper and validated against clinical data for a range of
Intensive care unit
patients lung abnormalities is embedded into the advisory system to predict continuously
Modelling
and non-invasively the patient’s respiratory response to changes in the ventilator settings.
Advisory system
The choice of appropriate ventilator settings involves finding a balance among a selection
Decision support system
of fundamentally competing therapeutic decisions. The design approach used here is based
Multi-objective optimisation
on a goal-directed multi-objective optimisation strategy to determine the optimal ventilator settings that effectively restore gas exchange and promote improved patient’s clinical conditions. As an initial step to its clinical validation, the advisory system’s closed-loop stability and performance have been assessed in a series of simulations scenarios reconstructed from real ICU patients data. The results show that the designed advisory system can generate good ventilator-setting advice under patient state changes and competing ventilator management targets. © 2010 Elsevier Ireland Ltd. All rights reserved.
∗
Corresponding author. Fax: +44 114 222 5607. E-mail addresses: [email protected] (A. Wang), m.mahfouf@sheffield.ac.uk (M. Mahfouf), [email protected] (G.H. Mills), g.panoutsos@sheffield.ac.uk (G. Panoutsos), d.linkens@sheffield.ac.uk (D.A. Linkens), [email protected] (K. Goode), [email protected] (H.-F. Kwok), m.denai@sheffield.ac.uk (M. Denaï). 0169-2607/$ – see front matter © 2010 Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.cmpb.2010.03.009
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 9 9 ( 2 0 1 0 ) 208–217
1.
Introduction
Mechanical ventilation is a-life-saving intervention which essentially aims to improve pulmonary gas exchange and reduce or take over the work of breathing in critically ill patients or patients with various forms of respiratory disorders. Respiratory failure can be caused either by a relative hypoventilation characterized by an increase in carbon dioxide (CO2 ) tension or failure of diffusion at the alveolar/endothelial interface (acute respiratory distress syndrome—ARDS, sepsis, etc.), characterized by decreased arterial oxygen (O2 ) tension. When initiating ventilation, the aim is to provide the patient with adequate tidal volume (VT ) delivered at a particular ventilatory/respiratory rate (RR). Oxygenation is improved by either raising the fraction of inspired oxygen (FiO2 ) or increasing the positive end-expiratory pressure (PEEP) in the case of severe hypoxia to prevent recruited lung units from collapsing at end-expiration. Another strategy is to prolong the inspiration time by increasing the ventilator’s inspiration-to-expiration (I:E) ratio. CO2 elimination is improved by increasing the minute volume (MV) either via an appropriate setting of VT or RR. The patient arterial blood gas (ABG) analysis, which includes the partial pressures of O2 (PaO2 ) and CO2 (PaCO2 ), has been the gold standard technique used to monitor the patient’s gas exchange status. The challenge for mechanical ventilation is to reduce both the duration of the patient’s respiratory support and the risk of ventilator-induced lung damage known as ventilator-induced lung injury (VILI). VILI can result from early complications caused by inadequate intubation or tracheotomy causing injuries and bleeding to the surrounding tissues of the oesophageous. Late complications often seen in ARDS patients and directly related to mechanical ventilation include barotraumas (repetitive closing and reopening of injured alveoli) caused by high airway pressures and/or volutrauma (alveolar overdistension when the lung units are physically stretched beyond their normal, maximum inflation point and/or alveolar disruption) caused by high tidal ventilation. Although mechanical ventilation has been used in intensive care units (ICU) for many years, selecting the best ventilation mode (or combination of modes) and particularly, adjusting the ventilator settings as the conditions and/or status of the patient change has remained a challenging task even for the most experienced clinicians. Standardised ventilator management protocols and guidelines used in most ICUs are disease-specific generalised approaches; not individually tailored for each patient’s existing pathophysiology. Decision support systems (DSS) are being recognised as potential tools for improving the quality of care in mechanically ventilated patients. They can assist clinicians interpret the monitored data and provide on-line patient-specific advice for setting ventilator parameters. A considerable amount of research has been conducted in this field and mainly three types of DSSs have emerged: (i) knowledge-based, (ii) model-based and (iii) combined knowledge- and model-based systems. Early knowledge-based systems were based on computerised protocols and guidelines acquired from the expert’s clinician and translated into computer interpretable set of
209
rules [1–3]. Their ultimate purpose was to ensure replicable clinical decisions for identical or equivalent patient’s states. More sophisticated collaborative knowledge-based systems using fuzzy reasoning and machine learning techniques to extract or discover knowledge from the monitored data have also been proposed [4]. Model-based techniques offer a more objectivist approach and have been successfully used in many engineering applications. However, in the field of ventilation management an accurate yet clinically useful patient’s physiological model is not easily obtainable. One successful model-based DSS for ventilation therapy was VentPlan [6]. The advice generated was based on a physiological model of the lungs [7,8] which describes O2 and CO2 gas exchange in the lungs and tissues and their transport around the circulation. VentPlan was evaluated by comparing the recommendations with changes in the ventilator settings recorded from surgical patients [6]. Recently, a comprehensive patient physiological model was developed [9] and used for optimising ventilator settings based on the evaluation of a utility function [5]. Very little research has been done in the development of combined knowledge-based and model-based DSSs for ventilation therapy. The system in Ref. [10] was designed to provide FiO2 , PEEP, RR and Pinsp (inspiratory pressure) settings for ventilated ICU patients. The advice includes a qualitative component obtained from a top-level module (knowledge-bases part) to indicate the direction of change of each ventilator setting and a quantitative component obtained from a lower level module (model-based part) to define the amount of changes for the relevant ventilator parameter. A recent review on different methodologies used in DSS for mechanical ventilation can be found in Ref. [11]. In Part 1 of this paper [12], a ventilated patient model SOPAVent has been developed and it can predict continuously and non-invasively the patient’s respiratory response to changes in ventilator settings. With the availability of the model and its proven prediction performance, this paves the way to develop an adaptive model-based decision support system for ventilator management. Model-based decision support system is more transparent and less subjective compared to a knowledge-based system. The users (ICU staff) can understand it more because it is based on a physiological model. The approach used in this paper is to use a goaldirected multi-objective optimisation strategy to determine the best ventilator settings to achieve the appropriate gas exchanges in the lungs. The paper is organised into five more sections. Section 2 gives an overview of the SOPAVent model; Section 3 presents the structure, design and implementation of the advisory system. Section 4 presents a simulation study to evaluate the performance of the advisory system in a series of simulated clinical scenarios. Section 5 discusses the current results and identifies the limitations. Finally, Section 6 summarises the conclusions drawn from this overall study.
2.
SOPAVent model
The model used in the development is named as SOPAVent (Simulation of Patient under Artificial Ventilation) and it was
210
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 9 9 ( 2 0 1 0 ) 208–217
• Shunt Shunt is estimated using a root-finding algorithm (secant algorithm) to match the model predicted PaO2 with the measured one. • Cardiac output (CO) CO is estimated using a population mean based on patient weight and height.
Fig. 1 – Schematic diagram of the SOPAVent model.
developed by our research group [13]. The SOPAVent model was developed using a set of mathematical equations representing the exchange of O2 and CO2 in the lungs and tissues together with their transport through the circulatory system based on respiratory physiology and mass balance equations. The model inputs are ventilator settings while the outputs are the arterial blood gases (PaO2 , PaCO2 ). In order to provide prediction, the model requires a set of parameters. The model structure is shown in Fig. 1. SOPAVent was implemented using MATLAB and SIMULINK. However, the application of the original model is limited due to the dependence of invasive measurements and the timeconsuming model tuning algorithm. Significant improvements have been made to improve the original SOPAVent model in order to improve its accuracy and applicability in routine ICU environment [10,14–17]. The improvements are mainly related to the model parameter estimations so that the model can become totally independent of any invasive measurements and generate good prediction performance for a wide range of ICU patients. The improvements on model parameter estimation are briefly summarised as follows: • Relative dead-space (Kd) An ANFIS model has been developed for Kd estimation. The model structure is as follows: Kd = F1 {PaCO2 , PEEP, Pinsp , RR, VT }, where F1 represents the ANFIS mapping. • Carbon dioxide production (VCO2 ) An ANFIS model has been developed for VCO2 estimation. The model structure is as follows: VCO2 = F2 {MV, VT , EtCO2 }, where F2 represents the ANFIS mapping. • Oxygen consumption (VO2 ) VO2 is estimated as the division of VCO2 and RQ (RQ is fixed as 0.8).
The improved SOPAVent model has become totally noninvasive. The model performance has been validated based on real patients’ data. Results show that the current SOPAVent model can generate good blood gas predictions responding to various combinations of ventilator setting changes and represent a wide range of ICU patients. SOPAVent has been validated to show that it can generate good steady-state predictions when patient’s conditions are relatively stable. However, in order to design a model-based decision support system that can allow for ventilator-setting advice continuously in a closed-loop mode, the patient model must be able to adapt itself to the patient state changes. Therefore, a continuously updated model structure has been developed based on SOPAVent to simulate the ventilated patients with their state evolution. The model structure is shown in Fig. 2. The continuously updated model was validated based on real patients’ data. The validation results show that, by frequently updating the five key model parameters (shunt, Kd, VCO2 , VO2 and raw) based on the continuous measurements from ICU, the model can represent the patient state accurately and lead to good blood gas predictions with the patient state evolution. The details of the SOPAVent model, its parameters model and its validation can be found in Part I of the paper [12]. The successful development of such a continuously updated patient model forms the basis for developing and evaluating an adaptive model-based decision support system for ventilator management. In the following sections, the design and evaluation of such a decision support system are presented.
3. Advisory system design and implementation As stated above, mechanical ventilation is a complex process aimed to provide adequate balance of oxygen levels (PaO2 ) and carbon dioxide (PaCO2 ) build up in the circulation. However, ventilation strategies that meet PaO2 and PaCO2 targets only are not always the optimal ones. Achieving desirable PaO2 and PaCO2 levels often requires an excessive airway pressure, VT and FiO2 , which could in turn harm the patient [18]. Therefore, any optimal ventilation strategy should represent a compromise between maintaining PaO2 and PaCO2 within acceptable ranges and minimizing the side effects of mechanical ventilation. Therefore, finding an optimal ventilation therapy specifically customised to a given patient’s medical conditions can be regarded as a goal-directed multiobjective optimisation problem. Using SOPAVent model, the patient response for different ventilator settings can be predicted. However, because the targets are competing, a compromise must be achieved. In this
211
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 9 9 ( 2 0 1 0 ) 208–217
Fig. 2 – Structure of the continuously updated SOPAVent(2) model.
Table 1 – The ventilator setting input range. Ventilator parameter
Input range
FiO2 (%) Pinsp (cm H2 O) Vrate (breath/min)
[30–100] [5–40] [4–20]
Table 2 – Control targets for mechanical ventilation in critically ill ICU patients. The normal range of parameters, as well as the clinically acceptable ranges. The variables may be allowed to go to some extent beyond the acceptable range if this is deemed clinically necessary. Controlled parameter
Normal range
PaO2 (kPa) PaCO2 (kPa) PIP (cm H2 O) VT (ml/body weight (kg))
11–13 5–6 ≤30 ≤7
Acceptable range 10–14 4–7 ≤35 ≤8
study, an aggregated multi-objective optimisation technique is applied. The method uses weight parameters to define the relative importance/prioritisation of each objective and sums the competing objectives into a single objective function which is then solved to determine the optimal ventilator parameters. The advisory system provides advice on FiO2 , Pinsp and Vrate settings only because of the unproven SOPAVent’s performance to predict PEEP and Tinsp . Because FiO2 mainly affects the patient’s oxygenation while Pinsp and Vrate mainly influence the minute volume ventilation, it was decided to divide the advisor into two subsystems (FiO2 and Pinsp /Vrate ). In order to generate a clinically meaningful ventilator-setting advice, the search intervals for FiO2 , Pinsp and Vrate have been set to the values given in Table 1. The control target for PaO2 , PaCO2 , PIP and VT as suggested by senior ICU clinicians are shown in Table 2. These values reveal that, in clinical practice, sub-optimal targets must be tolerated in order to reach a compromise among the many competing therapeutic goals.
3.1.
ing excessive airway pressure (PIP) and tidal volume (VT ). The objective function for the Pinsp /Vrate subsystem is formulated using the following aggregated multi-objective function:
J = 1
PaCO∗2 − PaCO2 PaCO2max
2
+ 2
V∗T − VT VTmax
2 + 3
PIPmax
(1)
where PaCO∗2 and V∗T denotes the targets for PaCO2 and VT respectively. PaCO2 and VT being SOPAVent actual predictions. The objective function J is subject to the following constraints: 5 ≤ Pinsp ≤ 40 and 4 ≤ Vrate ≤ 20; with PaCO∗2 = 5 kPa, V∗T = 7 ml/kg, VTmax = 1500 ml, PaCO2max = 20 kPa, PIP = Pinsp + PEEP, PIPmax = 50 cm H2 O and 1 , 2 , 3 ≥ 0. The objective function J consists of the weighted square sums of the normalised PaCO2 control, the weighted VT control error and a penalty function on PIP (Fig. 3). This penalty term on PIP is included to minimise the side effects caused by large airway pressures. The selection of the weighting parameters 1 , 2 and 3 is crucial as it decides on the relative importance of the individual goals and whether the optimal compromise among
Pinsp /Vrate subsystem
For this subsystem, the main goals are to maintain the patient’s PaCO2 within the assigned normal range while avoid-
PIP 6
Fig. 3 – PIP penalty function.
212
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 9 9 ( 2 0 1 0 ) 208–217
representing both the patient model embedded around the GA optimisation loop and the simulated patient model. The aim of the closed-loop validation is to assess the system’s ability to deal with different simulated patients’ scenarios designed to reproduce lung pathophysiological conditions similar to those observed in a real clinical environment and to evaluate whether the advisory system can produce clinically meaningful advice and consistent performance under various competing therapeutic goals. The clinical conditions of a patient with respiratory disorders treated with mechanical ventilation may deteriorate or improve over time. To simulate these clinical features, the following four scenarios have been considered:
Fig. 4 – Structure of ventilator management advisory system.
the competing therapeutic goals is achievable or not. In this project, the weight parameters 1 , 2 and 3 have been tuned based on the PaCO2 and VT control errors achieved and finally after a series of trial the values 1 = 0.4, 2 = 0.6 and 3 = 0 have been found appropriate to reach the desired compromise among the medical goals. The reason for setting 3 to 0 is that PIP constraint can be met systematically by VT control goals as they both limit the high Pinsp . Therefore, the PIP constraint is omitted in the objective function. Genetic algorithm (GA) [19] optimisation technique has been used in this study because of its proven performance in searching relatively large spaces without the need for information on function derivatives. The random combination of Pinsp and Vrate are searched and the values that generate the minimal objective function value are found.
3.2.
FiO2 subsystem
The main goal of the FiO2 subsystem is to maintain PaO2 within the normal range. The same method developed in Ref. [10] and applied here, consists of evaluating SOPAVent oxygen transport equations at steady state from which the first derivative of PaO2 to FiO2 can be derived. The Newton method is then used to search for the FiO2 in order to achieve the PaO2 target. The search for FiO2 to meet the PaO2 target can be described by the following equation: f (FiO2 ) = SOPAVent(FiO2 , ı) − PaO∗2 = 0
(2)
where ı stands for the model parameters and PaO∗2 = 12 kPa. The iterative formula for FiO2 is as follows: FiOn+1 = FiOn2 − 2
SOPAVent(FiOn2 , ı) − PaO∗2 (∂PaO2 /∂FiO2 )|FiOn
(3)
2
where FiOn2 denotes the nth successive approximation of FiO2 .
4. Evaluation of the advisory system in closed-loop The structure of the closed-loop simulation setup depicted in Fig. 4 includes the advisory system and the SOPAVent model
(1) An acute increase in shunt which then returns to the baseline level after approximately 1 h. (2) An acute increase in Kd which then returns to the baseline level after approximately 1 h. (3) A slow increase of relative dead-space (Kd). (4) A slow increase of shunt. Such scenarios are often encountered in the ICU, i.e. a dramatic increase in dead space would be to remove the blood supply to part of a lung as seen during pulmonary embolus. It is worth noting at this stage that much more frequently sudden changes will occur when a change of mode of ventilation occurs; perhaps the most extreme example would be going from spontaneous unsupported ventilation to positive pressure ventilation. This can have the effect of overdistending non-dependent airways and reducing their perfusion, whereas dependent airways would be compressed, but would still receive blood flow. The former is a paradigm that increases Kd, whereas the latter increases shunt. Once ventilated, if VT was reduced, the Kd would rise because Vd would increase relative to VT . The addition of additional volume in or inline with the catheter mount will increase dead space, such as is seen with the addition of a heat and moisture exchanger. The advisory system was designed to generate the ventilator-setting advice every 30 min. In the simulations, the parameters of the patient model SOPAVent were changed to reproduce the previous scenarios. Every 30 min, the simulated patient data (PaO2 , PaCO2 , VT , ventilator settings) were input to the advisory system to update the patient model. Thereafter, the predicted model outputs are passed to the GA-based multi-objective algorithm to calculate the ‘optimal’ ventilator settings. The derived ventilator settings were then input to the simulated patient to simulate the patient states in the next 30 min. Each simulation lasted 4 h and started with a 30 min period of stabilisation where the simulated patient’s ventilator settings were maintained at the initial values. A total of five (5) ICU patient data sets were used to reproduce the initial patient state in the simulated scenarios for the subsequent validation of the advisory system. The demographic information of the five patients are summarised in Table 3. The patients involved in the study represent a wide range of illness and severity (from mild to very severe) [15]. The results show that the advisory system was able to generate satisfactory and consistent performance for all the reconstructed patients’ scenarios. One patient’s results are presented in Figs. 5–8. This simulated scenario was recon-
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 9 9 ( 2 0 1 0 ) 208–217
Table 3 – The demographic information of the five ICU patients. Patient
Age
Gender
One Two Three Four Five
68 28 66 77 40
Male Female Male Female Male
Weight (kg) 57 68 68 57 71
Height (cm) 173 176 175 162 164
structed from a 68 years old male patient with weight = 57 kg and height = 173 cm. Fig. 5 shows how the patient state and ventilator settings change when Kd is increased from 0.37 to 0.58 over a period of 210 min. It can be seen that initially, the patient had a high PaCO2 at 8.26 kPa but a low VT at almost 6 ml/kg. The Vrate setting was at its maximum value of 20. After the initial 30 min stabilisation, the advisory system responded correctly by increasing the Pinsp from 13 cm H2 O to 17 cm H2 O. The advice on the Pinsp setting seems to have met the medical goals as it led to a reduction in PaCO2 within the acceptable range and without generating an excessive tidal volume and airway pressure. With the Kd increasing, PaCO2 would increase if the ventilator settings were unchanged. In this simulation, the advisor generated the best compromise among the competing goals by slightly raising Pinsp to respond to an increase in Kd. In the simulation results of Fig. 6, the shunt was slightly increased slightly from 11% to 22% over the 210 min period. As expected, an increase in shunt should reduce the PaO2 if the ventilator settings are kept unchanged and this is clearly illus-
213
trated in Fig. 6 by a slow decline of PaO2 between two sampling points. Every 30 min, the patient measurements were input to the advisor. The system responded correctly by increasing FiO2 and as a result PaO2 is improved and maintained close to the target level. For the other mechanical ventilation management goals, it can be seen that Pinsp was changed appropriately to maintain the conflicting targets, PaCO2 and VT , within their acceptable ranges. The simulation scenario of Fig. 7 corresponds to a severe change in the relative dead-space Kd from 0.37 to 0.57 during a period of 30 min and return to baseline after 1 h. This produced an increase in PaCO2 from 6.48 kPa to 8.21 kPa. In the current patient clinical condition, only Pinsp can be adjusted because the Vrate has reached its maximum setting. As a result, the advisor increased Pinsp from 17 cm H2 O to 19 cm H2 O to cause a reduction in PaCO2 from 8.21 kPa to 7.52 kPa. Although this also caused VT to increase from about 8 ml/kg to 9 ml/kg, in this particular case this response was deemed to be clinically acceptable. When the patient Kd was reduced from 0.52 to 0.37 1 h later, PaCO2 dropped from 7.52 kPa to 5.92 kPa, which was within the acceptable target range. The DSS responded correctly by reducing Pinsp to restore the VT to approximately 8 ml/kg. Fig. 8 shows how the patient state and ventilator settings changed when the patient shunt was increased from 11% to 17% in a 30 min period and returned to baseline after approximately 1 h. The increase in shunt led to the reduction of PaO2 from 12 kPa to 7.2 kPa. Thereafter, the advisor responded correctly by increasing FiO2 from 0.49 to 0.69 this led to a
Fig. 5 – Simulation of a slow increase in Kd with time.
214
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 9 9 ( 2 0 1 0 ) 208–217
Fig. 6 – Simulation of a slowly increasing shunt.
normalisation of PaO2 . When the shunt was reduced from 17% to 11% in a 30 min period 1 h later, the patient PaO2 was increased from 12 kPa to 27 kPa. The advisor responded correctly by reducing FiO2 from 0.69 to 0.48 in order to restore PaO2 back to its normal range. For the Pinsp /Vrate subsystem, the DSS adjusted Pinsp settings adequately to achieve a satisfactory compromise for both PaCO2 and VT control targets. From the closed-loop simulation results, it can be concluded that the advisor was able to generate satisfactory ventilator settings advice subject to the patient state changes and competing ventilation management targets. However, it is worth noting that the simulated patient scenarios may introduce a few problems. For example, the update of Kd in patient model cannot use the Kd data-driven model because the simulated patient data are far different from the real patient data which are used for Kd model training. Therefore, in this simulation study, the model estimated Kd and the simulated patient Kd were set to the same value. In addition to the qualitative approach of assessing the system’s performance, a quantitative assessment is conducted to statistically assess the ability of the decision support system to keep the blood gasses within the required target range. The percentage of time outside the management limits is used to indicate the performance of the decision support system. A “±10%” tolerance to the target values is added to the acceptable range defined in Table 2 so that the system evaluation can be more flexible and practical. In addition, the first 30-min simulation results are not included in the statistical analysis because it is regarded as a patient stabilisation period
(time needed for the patient blood gases to reach ‘steady state’). The five simulated patient simulation results are presented in Tables 4–7. It can be seen from the tables that the decision support system can generate good ventilator settings to manage the patient with the patient state evolution and when
Table 4 – Five patients, simulation performance with a slow increase of dead space. Patient
% Time out of range PaO2
One Two Three Four Five
0 0 0 0 0
PaCO2
VT
34.03 0 0 0 0
14.29 0 0 0 0
Table 5 – Five patients, simulation performance with a slow increase of shunt scenario. Patient
% Time out of range PaO2
One Two Three Four Five
0 0 0 0 0
PaCO2 0 0 0 0 0
VT 0 0 0 0 0
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 9 9 ( 2 0 1 0 ) 208–217
215
Fig. 7 – Simulation of a patient with acute Kd changes.
Table 6 – Five patients, simulation performance with an acute change of dead space. Patient
% Time out of range PaO2
One Two Three Four Five
0 0 0 0 0
PaCO2
VT
3.78 0 0 0 0
42.86 0 0 0 0
Table 7 – Five patients, simulation performance with an acute change of shunt. Patient
% Time out of range PaO2
One Two Three Four Five
17.84 24.59 21.20 16.70 9.64
PaCO2 0 0 0 0 0
VT 0 0 0 0 0
needed the system can find a balanced setting to satisfy the competing ventilation management multiple targets.
5.
Discussion
The proposed management target in Table 2 is a general guidance however, in practice, this is not hard limit and short-term
overshoot of these boundaries is acceptable in clinical practice. Often, the patient management targets are changed with the patient condition evolution. This will be included in future research to combine a flexible target-setting component into the DSS. The current evaluation is based on a simulation study. The simulated patient was initially constructed based on a real patient data but the evolving patient conditions are purely designed focusing on several simplified physiological parameter changes. Therefore, the current simulation study mainly demonstrates the approach and its optimisation algorithm performance. In future study, more efforts should be spent on the design of a simulated patient scenario based on real patients’ data stored in Patient Data Management System (PDMS). This will then represent a more physiological relevant simulated patient hence lead to more accurate and clinical relevant evaluations. As mentioned in Section 4, the evaluation study was originally designed to have two purposes: 1. To evaluate whether the advisory system can produce clinically meaningful advice and consistent performance under various competing therapeutic goals. 2. To evaluate whether the system can deal with different patients’ state evolution. The first goal has been achieved in this study. Because the current simulation results show the DSS adaptive ability by assuming the model updating is accurate, the adaptive feature of the DSS, therefore, has
216
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 9 9 ( 2 0 1 0 ) 208–217
Fig. 8 – Simulation of a patient with acute shunt changes.
not been fully validated. However, the continuous updated SOPAVent model performance has been validated based on real patients in Part I of the paper and it has shown a good performance [12]. A full validation of the DSS’s adaptive ability will be conducted with a more realistic simulated patient scenario in future study.
6.
Conclusions
A clinically useful advisory system for the management of patients under mechanical ventilation has been designed. Because ventilated patients are often critically ill and therefore clinically unstable, SOPAVent, a hybrid data-driven physiological model developed in a previous study was embedded into the advisor to predict continuously and non-invasively the patient’s respiratory response to changes in the ventilator settings. The state-of-the art approach used to derive the advisory system’s recommendations was based on the solution of an aggregated multi-objective/multi-criteria optimisation method. The method uses weighting parameters to define the relative importance/prioritisation of each objective and sums the competing objectives into a single objective function which is then solved using genetic algorithm to determine the optimal ventilator parameters. The advisory system has been validated using predefined scenarios with some physiological parameter changes. In all the simulation scenarios considered, the system was able to generate satisfactory ventilatory settings under competing therapeutic goals.
It remains for the future to evaluate the relevance, performance, efficiency and impact of the system in a real clinical setting. One of the main limitations of the current system is the unavailability of PEEP advice. Hence, one interesting area to investigate would be the extension of the current SOPAVent version to include a more accurate model of PEEP effects. In addition, the current advisory system is designed with fixed ventilator targets. However, in a clinical setting, these often vary depending on the patient’s condition. Hence, as part of the future work, a more flexible target-setting component should be included.
Conflicts of interest statement None declared.
Acknowledgements The authors wish to thank the anonymous reviewers for their comments which helped to improve the quality of this paper and they gratefully acknowledge the financial support of the UK Engineering and Physical Sciences Research Council (EPSRC) under Grant EP/520807/1.
references
[1] M. Dojat, F. Pachet, Z. Guessoum, D. Touchard, A. Harf, L. Brochard, NeoGanesh: a working system for the automated
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 9 9 ( 2 0 1 0 ) 208–217
[2]
[3]
[4]
[5]
[6]
[7] [8]
[9]
[10]
control of assisted ventilation in ICUs, Artificial Intelligence in Medicine 11 (2) (1997) 97–117. B.A. McKinley, F.A. Moore, R.M. Sailors, C.S. Cocanour, A. Marquez, R.K. Wright, A.S. Tonnesen, C.J. Wallace, A.H. Morris, T.D. East, Computerized decision support for mechanical ventilation of trauma induced ARDS: results of a randomized clinical trial, Journal of Trauma Injury Infection & Critical Care 50 (3) (2001) 415–424. T. Nemoto, G.E. Hatzakis, C.W. Thorpe, R. Olivenstein, S. Dial, J.H. Bates, Automatic control of pressure support mechanical ventilation using fuzzy logic, American Journal of Respiratory & Critical Care Medicine 160 (2) (1999) 550–556. S.Y. Belal, A.F.G. Taktak, A. Nevill, A. Spencer, An intelligent ventilation and oxygenation management system in neonatal intensive care using fuzzy trend template fitting, Physiological Measurement 26 (2005) 555–570. S.E. Rees, C. Allerod, D. Murley, Y. Zhao, B.W. Smith, S. Kjaergaard, P. Thorgaard, S. Andreassen, Using physiological models and decision theory for selecting appropriate ventilator settings, Journal of Clinical Monitoring & Computing 20 (6) (2006) 421–429. G.W. Rutledge, G.E. Thomsen, B.R. Farr, M.A. Tovar, J.X. Polaschek, I.A. Beinlich, L.B. Sheiner, L.M. Fagan, The design and implementation of a ventilator-management advisor, Artificial Intelligence in Medicine 5 (1) (1993) 67–82. C.J. Dickinson, A Computer Model of Human Respiration, MTP Press Limited, Lancaster, England, 1977. G. Thomsen, L. Sheiner, SIMV: an application of mathematical modelling in ventilator management, in: Proc. Thirteenth Annual Symposium on Computer Applications in Medical Care, Washington, DC, 1989, pp. 320–324. S. Andreassen, S.E. Rees, Mathematical models of oxygen and carbon dioxide storage and transport: interstitial fluid and tissue stores and whole-body transport, Critical Reviews in Biomedical Engineering 33 (3) (2005) 265–298. H.F. Kwok, D.A. Linkens, M. Mahfouf, G.H. Mills, SIVA: a hybrid knowledge-and model based advisory system for intensive care ventilators, IEEE Transactions on Information Technology in Biomedicine 8 (2) (2004) 161–172.
217
[11] F.T. Tehrani, J.H. Rourn, Intelligent decision support systems for mechanical ventilation, Artificial Intelligence in Medicine 44 (2008) 171–182. [12] M. Mahfouf, A. Wang, G.H. Mills, G. Panoutsos, D.A. Linkens, K. Goode, H.F. Kwok, M. Denai, Intelligent model-based advisory system for the management of ventilated intensive care patients. Hybrid blood gas patient model, Computer Methods and Programs in Biomedicine 99 (2010) 195–207. [13] K.M. Goode, Model-based development of a fuzzy logic advisor for artificially ventilated patients, Ph.D. Thesis, The University of Sheffield, 2001. [14] H.F. Kwok, SIVA: an intelligent advisory system for intensive care ventilators, Ph.D. Thesis, Department of Automatic Control and Systems Engineering, The University of Sheffield, Sheffield, UK, 2003. [15] A. Wang, Hybrid modelling and decision support for ventilator management in intensive care units, Ph.D. Thesis, Department of Automatic Control and Systems Engineering, The University of Sheffield, Sheffield, UK, 2008. [16] A. Wang, G. Panoutsos, M. Mahfouf, G.H. Mills, An improved blood gas intelligent hybrid model for mechanically ventilated patients in the Intensive Care Unit, in: IASTED BioMED’07, Proceedings of the 5th IASTED International Conference on Biomedical Engineering, February 14–16, 2007, Innsbruck, Austria, 2007, pp. 73–78. [17] A. Wang, G. Panoutsos, M. Mahfouf, M. Mills, Goal-directed therapy for general ICU patients using aggregated multi-objective optimisation, in: 17th IFAC World Congress, 6–11 July, 2008, Seoul, Korea, 2008. [18] ARDSnetwork, Ventilation with lower tidal volumes as compared with traditional tidal volumes for acute lung injury and the acute respiratory distress syndrome, The Acute Respiratory Distress Syndrome Network, New England Journal of Medicine 342 (18) (2000) 1301– 1308. [19] D.E. Goldberg, Genetic Algorithms in Search, in: Optimization and Machine Learning, Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA, 1989.
journal homepage: www.intl.elsevierhealth.com/journals/cmpb
Intelligent model-based advisory system for the management of ventilated intensive care patients. Part II: Advisory system design and evaluation Ang Wang a , Mahdi Mahfouf b,∗ , Gary H. Mills c , G. Panoutsos b , D.A. Linkens b , K. Goode d , Hoi-Fei Kwok e , Mouloud Denaï a a
Process Automation, ABB Limited, Howard Road, Eaton Socon, Cambridgeshire PE19 8EU, UK Department of Automatic and Systems Engineering, The University of Sheffield, Mappin Street, Sheffield S1 3JD, UK c Department of Critical Care and Anaesthesia, Northern General Hospital, Herries Road, Sheffield S5 7AU, UK d Postgraduate Medical Institute, The University of Hull, Cottingham Road, Hull HU6 7RX, UK e School of Psychology, The University of Birmingham, Edgbaston, Birmingham B15 2TT, UK b
a r t i c l e
i n f o
a b s t r a c t
Article history:
The optimisation of ventilatory support is a crucial issue for the management of respira-
Received 10 April 2009
tory failure in critically ill patients, aiming at improving gas exchange while preventing
Received in revised form
ventilator-induced dysfunction of the respiratory system. Clinicians often rely on their
15 March 2010
knowledge/experience and regular observation of the patient’s response for adjusting the
Accepted 15 March 2010
level of respiratory support. Using a similar data-driven decision-making methodology, an adaptive model-based advisory system has been designed for the clinical monitoring
Keywords:
and management of mechanically ventilated patients. The hybrid blood gas patient model
Mechanical ventilation
SOPAVent developed in Part I of this paper and validated against clinical data for a range of
Intensive care unit
patients lung abnormalities is embedded into the advisory system to predict continuously
Modelling
and non-invasively the patient’s respiratory response to changes in the ventilator settings.
Advisory system
The choice of appropriate ventilator settings involves finding a balance among a selection
Decision support system
of fundamentally competing therapeutic decisions. The design approach used here is based
Multi-objective optimisation
on a goal-directed multi-objective optimisation strategy to determine the optimal ventilator settings that effectively restore gas exchange and promote improved patient’s clinical conditions. As an initial step to its clinical validation, the advisory system’s closed-loop stability and performance have been assessed in a series of simulations scenarios reconstructed from real ICU patients data. The results show that the designed advisory system can generate good ventilator-setting advice under patient state changes and competing ventilator management targets. © 2010 Elsevier Ireland Ltd. All rights reserved.
∗
Corresponding author. Fax: +44 114 222 5607. E-mail addresses: [email protected] (A. Wang), m.mahfouf@sheffield.ac.uk (M. Mahfouf), [email protected] (G.H. Mills), g.panoutsos@sheffield.ac.uk (G. Panoutsos), d.linkens@sheffield.ac.uk (D.A. Linkens), [email protected] (K. Goode), [email protected] (H.-F. Kwok), m.denai@sheffield.ac.uk (M. Denaï). 0169-2607/$ – see front matter © 2010 Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.cmpb.2010.03.009
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 9 9 ( 2 0 1 0 ) 208–217
1.
Introduction
Mechanical ventilation is a-life-saving intervention which essentially aims to improve pulmonary gas exchange and reduce or take over the work of breathing in critically ill patients or patients with various forms of respiratory disorders. Respiratory failure can be caused either by a relative hypoventilation characterized by an increase in carbon dioxide (CO2 ) tension or failure of diffusion at the alveolar/endothelial interface (acute respiratory distress syndrome—ARDS, sepsis, etc.), characterized by decreased arterial oxygen (O2 ) tension. When initiating ventilation, the aim is to provide the patient with adequate tidal volume (VT ) delivered at a particular ventilatory/respiratory rate (RR). Oxygenation is improved by either raising the fraction of inspired oxygen (FiO2 ) or increasing the positive end-expiratory pressure (PEEP) in the case of severe hypoxia to prevent recruited lung units from collapsing at end-expiration. Another strategy is to prolong the inspiration time by increasing the ventilator’s inspiration-to-expiration (I:E) ratio. CO2 elimination is improved by increasing the minute volume (MV) either via an appropriate setting of VT or RR. The patient arterial blood gas (ABG) analysis, which includes the partial pressures of O2 (PaO2 ) and CO2 (PaCO2 ), has been the gold standard technique used to monitor the patient’s gas exchange status. The challenge for mechanical ventilation is to reduce both the duration of the patient’s respiratory support and the risk of ventilator-induced lung damage known as ventilator-induced lung injury (VILI). VILI can result from early complications caused by inadequate intubation or tracheotomy causing injuries and bleeding to the surrounding tissues of the oesophageous. Late complications often seen in ARDS patients and directly related to mechanical ventilation include barotraumas (repetitive closing and reopening of injured alveoli) caused by high airway pressures and/or volutrauma (alveolar overdistension when the lung units are physically stretched beyond their normal, maximum inflation point and/or alveolar disruption) caused by high tidal ventilation. Although mechanical ventilation has been used in intensive care units (ICU) for many years, selecting the best ventilation mode (or combination of modes) and particularly, adjusting the ventilator settings as the conditions and/or status of the patient change has remained a challenging task even for the most experienced clinicians. Standardised ventilator management protocols and guidelines used in most ICUs are disease-specific generalised approaches; not individually tailored for each patient’s existing pathophysiology. Decision support systems (DSS) are being recognised as potential tools for improving the quality of care in mechanically ventilated patients. They can assist clinicians interpret the monitored data and provide on-line patient-specific advice for setting ventilator parameters. A considerable amount of research has been conducted in this field and mainly three types of DSSs have emerged: (i) knowledge-based, (ii) model-based and (iii) combined knowledge- and model-based systems. Early knowledge-based systems were based on computerised protocols and guidelines acquired from the expert’s clinician and translated into computer interpretable set of
209
rules [1–3]. Their ultimate purpose was to ensure replicable clinical decisions for identical or equivalent patient’s states. More sophisticated collaborative knowledge-based systems using fuzzy reasoning and machine learning techniques to extract or discover knowledge from the monitored data have also been proposed [4]. Model-based techniques offer a more objectivist approach and have been successfully used in many engineering applications. However, in the field of ventilation management an accurate yet clinically useful patient’s physiological model is not easily obtainable. One successful model-based DSS for ventilation therapy was VentPlan [6]. The advice generated was based on a physiological model of the lungs [7,8] which describes O2 and CO2 gas exchange in the lungs and tissues and their transport around the circulation. VentPlan was evaluated by comparing the recommendations with changes in the ventilator settings recorded from surgical patients [6]. Recently, a comprehensive patient physiological model was developed [9] and used for optimising ventilator settings based on the evaluation of a utility function [5]. Very little research has been done in the development of combined knowledge-based and model-based DSSs for ventilation therapy. The system in Ref. [10] was designed to provide FiO2 , PEEP, RR and Pinsp (inspiratory pressure) settings for ventilated ICU patients. The advice includes a qualitative component obtained from a top-level module (knowledge-bases part) to indicate the direction of change of each ventilator setting and a quantitative component obtained from a lower level module (model-based part) to define the amount of changes for the relevant ventilator parameter. A recent review on different methodologies used in DSS for mechanical ventilation can be found in Ref. [11]. In Part 1 of this paper [12], a ventilated patient model SOPAVent has been developed and it can predict continuously and non-invasively the patient’s respiratory response to changes in ventilator settings. With the availability of the model and its proven prediction performance, this paves the way to develop an adaptive model-based decision support system for ventilator management. Model-based decision support system is more transparent and less subjective compared to a knowledge-based system. The users (ICU staff) can understand it more because it is based on a physiological model. The approach used in this paper is to use a goaldirected multi-objective optimisation strategy to determine the best ventilator settings to achieve the appropriate gas exchanges in the lungs. The paper is organised into five more sections. Section 2 gives an overview of the SOPAVent model; Section 3 presents the structure, design and implementation of the advisory system. Section 4 presents a simulation study to evaluate the performance of the advisory system in a series of simulated clinical scenarios. Section 5 discusses the current results and identifies the limitations. Finally, Section 6 summarises the conclusions drawn from this overall study.
2.
SOPAVent model
The model used in the development is named as SOPAVent (Simulation of Patient under Artificial Ventilation) and it was
210
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 9 9 ( 2 0 1 0 ) 208–217
• Shunt Shunt is estimated using a root-finding algorithm (secant algorithm) to match the model predicted PaO2 with the measured one. • Cardiac output (CO) CO is estimated using a population mean based on patient weight and height.
Fig. 1 – Schematic diagram of the SOPAVent model.
developed by our research group [13]. The SOPAVent model was developed using a set of mathematical equations representing the exchange of O2 and CO2 in the lungs and tissues together with their transport through the circulatory system based on respiratory physiology and mass balance equations. The model inputs are ventilator settings while the outputs are the arterial blood gases (PaO2 , PaCO2 ). In order to provide prediction, the model requires a set of parameters. The model structure is shown in Fig. 1. SOPAVent was implemented using MATLAB and SIMULINK. However, the application of the original model is limited due to the dependence of invasive measurements and the timeconsuming model tuning algorithm. Significant improvements have been made to improve the original SOPAVent model in order to improve its accuracy and applicability in routine ICU environment [10,14–17]. The improvements are mainly related to the model parameter estimations so that the model can become totally independent of any invasive measurements and generate good prediction performance for a wide range of ICU patients. The improvements on model parameter estimation are briefly summarised as follows: • Relative dead-space (Kd) An ANFIS model has been developed for Kd estimation. The model structure is as follows: Kd = F1 {PaCO2 , PEEP, Pinsp , RR, VT }, where F1 represents the ANFIS mapping. • Carbon dioxide production (VCO2 ) An ANFIS model has been developed for VCO2 estimation. The model structure is as follows: VCO2 = F2 {MV, VT , EtCO2 }, where F2 represents the ANFIS mapping. • Oxygen consumption (VO2 ) VO2 is estimated as the division of VCO2 and RQ (RQ is fixed as 0.8).
The improved SOPAVent model has become totally noninvasive. The model performance has been validated based on real patients’ data. Results show that the current SOPAVent model can generate good blood gas predictions responding to various combinations of ventilator setting changes and represent a wide range of ICU patients. SOPAVent has been validated to show that it can generate good steady-state predictions when patient’s conditions are relatively stable. However, in order to design a model-based decision support system that can allow for ventilator-setting advice continuously in a closed-loop mode, the patient model must be able to adapt itself to the patient state changes. Therefore, a continuously updated model structure has been developed based on SOPAVent to simulate the ventilated patients with their state evolution. The model structure is shown in Fig. 2. The continuously updated model was validated based on real patients’ data. The validation results show that, by frequently updating the five key model parameters (shunt, Kd, VCO2 , VO2 and raw) based on the continuous measurements from ICU, the model can represent the patient state accurately and lead to good blood gas predictions with the patient state evolution. The details of the SOPAVent model, its parameters model and its validation can be found in Part I of the paper [12]. The successful development of such a continuously updated patient model forms the basis for developing and evaluating an adaptive model-based decision support system for ventilator management. In the following sections, the design and evaluation of such a decision support system are presented.
3. Advisory system design and implementation As stated above, mechanical ventilation is a complex process aimed to provide adequate balance of oxygen levels (PaO2 ) and carbon dioxide (PaCO2 ) build up in the circulation. However, ventilation strategies that meet PaO2 and PaCO2 targets only are not always the optimal ones. Achieving desirable PaO2 and PaCO2 levels often requires an excessive airway pressure, VT and FiO2 , which could in turn harm the patient [18]. Therefore, any optimal ventilation strategy should represent a compromise between maintaining PaO2 and PaCO2 within acceptable ranges and minimizing the side effects of mechanical ventilation. Therefore, finding an optimal ventilation therapy specifically customised to a given patient’s medical conditions can be regarded as a goal-directed multiobjective optimisation problem. Using SOPAVent model, the patient response for different ventilator settings can be predicted. However, because the targets are competing, a compromise must be achieved. In this
211
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 9 9 ( 2 0 1 0 ) 208–217
Fig. 2 – Structure of the continuously updated SOPAVent(2) model.
Table 1 – The ventilator setting input range. Ventilator parameter
Input range
FiO2 (%) Pinsp (cm H2 O) Vrate (breath/min)
[30–100] [5–40] [4–20]
Table 2 – Control targets for mechanical ventilation in critically ill ICU patients. The normal range of parameters, as well as the clinically acceptable ranges. The variables may be allowed to go to some extent beyond the acceptable range if this is deemed clinically necessary. Controlled parameter
Normal range
PaO2 (kPa) PaCO2 (kPa) PIP (cm H2 O) VT (ml/body weight (kg))
11–13 5–6 ≤30 ≤7
Acceptable range 10–14 4–7 ≤35 ≤8
study, an aggregated multi-objective optimisation technique is applied. The method uses weight parameters to define the relative importance/prioritisation of each objective and sums the competing objectives into a single objective function which is then solved to determine the optimal ventilator parameters. The advisory system provides advice on FiO2 , Pinsp and Vrate settings only because of the unproven SOPAVent’s performance to predict PEEP and Tinsp . Because FiO2 mainly affects the patient’s oxygenation while Pinsp and Vrate mainly influence the minute volume ventilation, it was decided to divide the advisor into two subsystems (FiO2 and Pinsp /Vrate ). In order to generate a clinically meaningful ventilator-setting advice, the search intervals for FiO2 , Pinsp and Vrate have been set to the values given in Table 1. The control target for PaO2 , PaCO2 , PIP and VT as suggested by senior ICU clinicians are shown in Table 2. These values reveal that, in clinical practice, sub-optimal targets must be tolerated in order to reach a compromise among the many competing therapeutic goals.
3.1.
ing excessive airway pressure (PIP) and tidal volume (VT ). The objective function for the Pinsp /Vrate subsystem is formulated using the following aggregated multi-objective function:
J = 1
PaCO∗2 − PaCO2 PaCO2max
2
+ 2
V∗T − VT VTmax
2 + 3
PIPmax
(1)
where PaCO∗2 and V∗T denotes the targets for PaCO2 and VT respectively. PaCO2 and VT being SOPAVent actual predictions. The objective function J is subject to the following constraints: 5 ≤ Pinsp ≤ 40 and 4 ≤ Vrate ≤ 20; with PaCO∗2 = 5 kPa, V∗T = 7 ml/kg, VTmax = 1500 ml, PaCO2max = 20 kPa, PIP = Pinsp + PEEP, PIPmax = 50 cm H2 O and 1 , 2 , 3 ≥ 0. The objective function J consists of the weighted square sums of the normalised PaCO2 control, the weighted VT control error and a penalty function on PIP (Fig. 3). This penalty term on PIP is included to minimise the side effects caused by large airway pressures. The selection of the weighting parameters 1 , 2 and 3 is crucial as it decides on the relative importance of the individual goals and whether the optimal compromise among
Pinsp /Vrate subsystem
For this subsystem, the main goals are to maintain the patient’s PaCO2 within the assigned normal range while avoid-
PIP 6
Fig. 3 – PIP penalty function.
212
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 9 9 ( 2 0 1 0 ) 208–217
representing both the patient model embedded around the GA optimisation loop and the simulated patient model. The aim of the closed-loop validation is to assess the system’s ability to deal with different simulated patients’ scenarios designed to reproduce lung pathophysiological conditions similar to those observed in a real clinical environment and to evaluate whether the advisory system can produce clinically meaningful advice and consistent performance under various competing therapeutic goals. The clinical conditions of a patient with respiratory disorders treated with mechanical ventilation may deteriorate or improve over time. To simulate these clinical features, the following four scenarios have been considered:
Fig. 4 – Structure of ventilator management advisory system.
the competing therapeutic goals is achievable or not. In this project, the weight parameters 1 , 2 and 3 have been tuned based on the PaCO2 and VT control errors achieved and finally after a series of trial the values 1 = 0.4, 2 = 0.6 and 3 = 0 have been found appropriate to reach the desired compromise among the medical goals. The reason for setting 3 to 0 is that PIP constraint can be met systematically by VT control goals as they both limit the high Pinsp . Therefore, the PIP constraint is omitted in the objective function. Genetic algorithm (GA) [19] optimisation technique has been used in this study because of its proven performance in searching relatively large spaces without the need for information on function derivatives. The random combination of Pinsp and Vrate are searched and the values that generate the minimal objective function value are found.
3.2.
FiO2 subsystem
The main goal of the FiO2 subsystem is to maintain PaO2 within the normal range. The same method developed in Ref. [10] and applied here, consists of evaluating SOPAVent oxygen transport equations at steady state from which the first derivative of PaO2 to FiO2 can be derived. The Newton method is then used to search for the FiO2 in order to achieve the PaO2 target. The search for FiO2 to meet the PaO2 target can be described by the following equation: f (FiO2 ) = SOPAVent(FiO2 , ı) − PaO∗2 = 0
(2)
where ı stands for the model parameters and PaO∗2 = 12 kPa. The iterative formula for FiO2 is as follows: FiOn+1 = FiOn2 − 2
SOPAVent(FiOn2 , ı) − PaO∗2 (∂PaO2 /∂FiO2 )|FiOn
(3)
2
where FiOn2 denotes the nth successive approximation of FiO2 .
4. Evaluation of the advisory system in closed-loop The structure of the closed-loop simulation setup depicted in Fig. 4 includes the advisory system and the SOPAVent model
(1) An acute increase in shunt which then returns to the baseline level after approximately 1 h. (2) An acute increase in Kd which then returns to the baseline level after approximately 1 h. (3) A slow increase of relative dead-space (Kd). (4) A slow increase of shunt. Such scenarios are often encountered in the ICU, i.e. a dramatic increase in dead space would be to remove the blood supply to part of a lung as seen during pulmonary embolus. It is worth noting at this stage that much more frequently sudden changes will occur when a change of mode of ventilation occurs; perhaps the most extreme example would be going from spontaneous unsupported ventilation to positive pressure ventilation. This can have the effect of overdistending non-dependent airways and reducing their perfusion, whereas dependent airways would be compressed, but would still receive blood flow. The former is a paradigm that increases Kd, whereas the latter increases shunt. Once ventilated, if VT was reduced, the Kd would rise because Vd would increase relative to VT . The addition of additional volume in or inline with the catheter mount will increase dead space, such as is seen with the addition of a heat and moisture exchanger. The advisory system was designed to generate the ventilator-setting advice every 30 min. In the simulations, the parameters of the patient model SOPAVent were changed to reproduce the previous scenarios. Every 30 min, the simulated patient data (PaO2 , PaCO2 , VT , ventilator settings) were input to the advisory system to update the patient model. Thereafter, the predicted model outputs are passed to the GA-based multi-objective algorithm to calculate the ‘optimal’ ventilator settings. The derived ventilator settings were then input to the simulated patient to simulate the patient states in the next 30 min. Each simulation lasted 4 h and started with a 30 min period of stabilisation where the simulated patient’s ventilator settings were maintained at the initial values. A total of five (5) ICU patient data sets were used to reproduce the initial patient state in the simulated scenarios for the subsequent validation of the advisory system. The demographic information of the five patients are summarised in Table 3. The patients involved in the study represent a wide range of illness and severity (from mild to very severe) [15]. The results show that the advisory system was able to generate satisfactory and consistent performance for all the reconstructed patients’ scenarios. One patient’s results are presented in Figs. 5–8. This simulated scenario was recon-
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 9 9 ( 2 0 1 0 ) 208–217
Table 3 – The demographic information of the five ICU patients. Patient
Age
Gender
One Two Three Four Five
68 28 66 77 40
Male Female Male Female Male
Weight (kg) 57 68 68 57 71
Height (cm) 173 176 175 162 164
structed from a 68 years old male patient with weight = 57 kg and height = 173 cm. Fig. 5 shows how the patient state and ventilator settings change when Kd is increased from 0.37 to 0.58 over a period of 210 min. It can be seen that initially, the patient had a high PaCO2 at 8.26 kPa but a low VT at almost 6 ml/kg. The Vrate setting was at its maximum value of 20. After the initial 30 min stabilisation, the advisory system responded correctly by increasing the Pinsp from 13 cm H2 O to 17 cm H2 O. The advice on the Pinsp setting seems to have met the medical goals as it led to a reduction in PaCO2 within the acceptable range and without generating an excessive tidal volume and airway pressure. With the Kd increasing, PaCO2 would increase if the ventilator settings were unchanged. In this simulation, the advisor generated the best compromise among the competing goals by slightly raising Pinsp to respond to an increase in Kd. In the simulation results of Fig. 6, the shunt was slightly increased slightly from 11% to 22% over the 210 min period. As expected, an increase in shunt should reduce the PaO2 if the ventilator settings are kept unchanged and this is clearly illus-
213
trated in Fig. 6 by a slow decline of PaO2 between two sampling points. Every 30 min, the patient measurements were input to the advisor. The system responded correctly by increasing FiO2 and as a result PaO2 is improved and maintained close to the target level. For the other mechanical ventilation management goals, it can be seen that Pinsp was changed appropriately to maintain the conflicting targets, PaCO2 and VT , within their acceptable ranges. The simulation scenario of Fig. 7 corresponds to a severe change in the relative dead-space Kd from 0.37 to 0.57 during a period of 30 min and return to baseline after 1 h. This produced an increase in PaCO2 from 6.48 kPa to 8.21 kPa. In the current patient clinical condition, only Pinsp can be adjusted because the Vrate has reached its maximum setting. As a result, the advisor increased Pinsp from 17 cm H2 O to 19 cm H2 O to cause a reduction in PaCO2 from 8.21 kPa to 7.52 kPa. Although this also caused VT to increase from about 8 ml/kg to 9 ml/kg, in this particular case this response was deemed to be clinically acceptable. When the patient Kd was reduced from 0.52 to 0.37 1 h later, PaCO2 dropped from 7.52 kPa to 5.92 kPa, which was within the acceptable target range. The DSS responded correctly by reducing Pinsp to restore the VT to approximately 8 ml/kg. Fig. 8 shows how the patient state and ventilator settings changed when the patient shunt was increased from 11% to 17% in a 30 min period and returned to baseline after approximately 1 h. The increase in shunt led to the reduction of PaO2 from 12 kPa to 7.2 kPa. Thereafter, the advisor responded correctly by increasing FiO2 from 0.49 to 0.69 this led to a
Fig. 5 – Simulation of a slow increase in Kd with time.
214
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 9 9 ( 2 0 1 0 ) 208–217
Fig. 6 – Simulation of a slowly increasing shunt.
normalisation of PaO2 . When the shunt was reduced from 17% to 11% in a 30 min period 1 h later, the patient PaO2 was increased from 12 kPa to 27 kPa. The advisor responded correctly by reducing FiO2 from 0.69 to 0.48 in order to restore PaO2 back to its normal range. For the Pinsp /Vrate subsystem, the DSS adjusted Pinsp settings adequately to achieve a satisfactory compromise for both PaCO2 and VT control targets. From the closed-loop simulation results, it can be concluded that the advisor was able to generate satisfactory ventilator settings advice subject to the patient state changes and competing ventilation management targets. However, it is worth noting that the simulated patient scenarios may introduce a few problems. For example, the update of Kd in patient model cannot use the Kd data-driven model because the simulated patient data are far different from the real patient data which are used for Kd model training. Therefore, in this simulation study, the model estimated Kd and the simulated patient Kd were set to the same value. In addition to the qualitative approach of assessing the system’s performance, a quantitative assessment is conducted to statistically assess the ability of the decision support system to keep the blood gasses within the required target range. The percentage of time outside the management limits is used to indicate the performance of the decision support system. A “±10%” tolerance to the target values is added to the acceptable range defined in Table 2 so that the system evaluation can be more flexible and practical. In addition, the first 30-min simulation results are not included in the statistical analysis because it is regarded as a patient stabilisation period
(time needed for the patient blood gases to reach ‘steady state’). The five simulated patient simulation results are presented in Tables 4–7. It can be seen from the tables that the decision support system can generate good ventilator settings to manage the patient with the patient state evolution and when
Table 4 – Five patients, simulation performance with a slow increase of dead space. Patient
% Time out of range PaO2
One Two Three Four Five
0 0 0 0 0
PaCO2
VT
34.03 0 0 0 0
14.29 0 0 0 0
Table 5 – Five patients, simulation performance with a slow increase of shunt scenario. Patient
% Time out of range PaO2
One Two Three Four Five
0 0 0 0 0
PaCO2 0 0 0 0 0
VT 0 0 0 0 0
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 9 9 ( 2 0 1 0 ) 208–217
215
Fig. 7 – Simulation of a patient with acute Kd changes.
Table 6 – Five patients, simulation performance with an acute change of dead space. Patient
% Time out of range PaO2
One Two Three Four Five
0 0 0 0 0
PaCO2
VT
3.78 0 0 0 0
42.86 0 0 0 0
Table 7 – Five patients, simulation performance with an acute change of shunt. Patient
% Time out of range PaO2
One Two Three Four Five
17.84 24.59 21.20 16.70 9.64
PaCO2 0 0 0 0 0
VT 0 0 0 0 0
needed the system can find a balanced setting to satisfy the competing ventilation management multiple targets.
5.
Discussion
The proposed management target in Table 2 is a general guidance however, in practice, this is not hard limit and short-term
overshoot of these boundaries is acceptable in clinical practice. Often, the patient management targets are changed with the patient condition evolution. This will be included in future research to combine a flexible target-setting component into the DSS. The current evaluation is based on a simulation study. The simulated patient was initially constructed based on a real patient data but the evolving patient conditions are purely designed focusing on several simplified physiological parameter changes. Therefore, the current simulation study mainly demonstrates the approach and its optimisation algorithm performance. In future study, more efforts should be spent on the design of a simulated patient scenario based on real patients’ data stored in Patient Data Management System (PDMS). This will then represent a more physiological relevant simulated patient hence lead to more accurate and clinical relevant evaluations. As mentioned in Section 4, the evaluation study was originally designed to have two purposes: 1. To evaluate whether the advisory system can produce clinically meaningful advice and consistent performance under various competing therapeutic goals. 2. To evaluate whether the system can deal with different patients’ state evolution. The first goal has been achieved in this study. Because the current simulation results show the DSS adaptive ability by assuming the model updating is accurate, the adaptive feature of the DSS, therefore, has
216
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 9 9 ( 2 0 1 0 ) 208–217
Fig. 8 – Simulation of a patient with acute shunt changes.
not been fully validated. However, the continuous updated SOPAVent model performance has been validated based on real patients in Part I of the paper and it has shown a good performance [12]. A full validation of the DSS’s adaptive ability will be conducted with a more realistic simulated patient scenario in future study.
6.
Conclusions
A clinically useful advisory system for the management of patients under mechanical ventilation has been designed. Because ventilated patients are often critically ill and therefore clinically unstable, SOPAVent, a hybrid data-driven physiological model developed in a previous study was embedded into the advisor to predict continuously and non-invasively the patient’s respiratory response to changes in the ventilator settings. The state-of-the art approach used to derive the advisory system’s recommendations was based on the solution of an aggregated multi-objective/multi-criteria optimisation method. The method uses weighting parameters to define the relative importance/prioritisation of each objective and sums the competing objectives into a single objective function which is then solved using genetic algorithm to determine the optimal ventilator parameters. The advisory system has been validated using predefined scenarios with some physiological parameter changes. In all the simulation scenarios considered, the system was able to generate satisfactory ventilatory settings under competing therapeutic goals.
It remains for the future to evaluate the relevance, performance, efficiency and impact of the system in a real clinical setting. One of the main limitations of the current system is the unavailability of PEEP advice. Hence, one interesting area to investigate would be the extension of the current SOPAVent version to include a more accurate model of PEEP effects. In addition, the current advisory system is designed with fixed ventilator targets. However, in a clinical setting, these often vary depending on the patient’s condition. Hence, as part of the future work, a more flexible target-setting component should be included.
Conflicts of interest statement None declared.
Acknowledgements The authors wish to thank the anonymous reviewers for their comments which helped to improve the quality of this paper and they gratefully acknowledge the financial support of the UK Engineering and Physical Sciences Research Council (EPSRC) under Grant EP/520807/1.
references
[1] M. Dojat, F. Pachet, Z. Guessoum, D. Touchard, A. Harf, L. Brochard, NeoGanesh: a working system for the automated
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 9 9 ( 2 0 1 0 ) 208–217
[2]
[3]
[4]
[5]
[6]
[7] [8]
[9]
[10]
control of assisted ventilation in ICUs, Artificial Intelligence in Medicine 11 (2) (1997) 97–117. B.A. McKinley, F.A. Moore, R.M. Sailors, C.S. Cocanour, A. Marquez, R.K. Wright, A.S. Tonnesen, C.J. Wallace, A.H. Morris, T.D. East, Computerized decision support for mechanical ventilation of trauma induced ARDS: results of a randomized clinical trial, Journal of Trauma Injury Infection & Critical Care 50 (3) (2001) 415–424. T. Nemoto, G.E. Hatzakis, C.W. Thorpe, R. Olivenstein, S. Dial, J.H. Bates, Automatic control of pressure support mechanical ventilation using fuzzy logic, American Journal of Respiratory & Critical Care Medicine 160 (2) (1999) 550–556. S.Y. Belal, A.F.G. Taktak, A. Nevill, A. Spencer, An intelligent ventilation and oxygenation management system in neonatal intensive care using fuzzy trend template fitting, Physiological Measurement 26 (2005) 555–570. S.E. Rees, C. Allerod, D. Murley, Y. Zhao, B.W. Smith, S. Kjaergaard, P. Thorgaard, S. Andreassen, Using physiological models and decision theory for selecting appropriate ventilator settings, Journal of Clinical Monitoring & Computing 20 (6) (2006) 421–429. G.W. Rutledge, G.E. Thomsen, B.R. Farr, M.A. Tovar, J.X. Polaschek, I.A. Beinlich, L.B. Sheiner, L.M. Fagan, The design and implementation of a ventilator-management advisor, Artificial Intelligence in Medicine 5 (1) (1993) 67–82. C.J. Dickinson, A Computer Model of Human Respiration, MTP Press Limited, Lancaster, England, 1977. G. Thomsen, L. Sheiner, SIMV: an application of mathematical modelling in ventilator management, in: Proc. Thirteenth Annual Symposium on Computer Applications in Medical Care, Washington, DC, 1989, pp. 320–324. S. Andreassen, S.E. Rees, Mathematical models of oxygen and carbon dioxide storage and transport: interstitial fluid and tissue stores and whole-body transport, Critical Reviews in Biomedical Engineering 33 (3) (2005) 265–298. H.F. Kwok, D.A. Linkens, M. Mahfouf, G.H. Mills, SIVA: a hybrid knowledge-and model based advisory system for intensive care ventilators, IEEE Transactions on Information Technology in Biomedicine 8 (2) (2004) 161–172.
217
[11] F.T. Tehrani, J.H. Rourn, Intelligent decision support systems for mechanical ventilation, Artificial Intelligence in Medicine 44 (2008) 171–182. [12] M. Mahfouf, A. Wang, G.H. Mills, G. Panoutsos, D.A. Linkens, K. Goode, H.F. Kwok, M. Denai, Intelligent model-based advisory system for the management of ventilated intensive care patients. Hybrid blood gas patient model, Computer Methods and Programs in Biomedicine 99 (2010) 195–207. [13] K.M. Goode, Model-based development of a fuzzy logic advisor for artificially ventilated patients, Ph.D. Thesis, The University of Sheffield, 2001. [14] H.F. Kwok, SIVA: an intelligent advisory system for intensive care ventilators, Ph.D. Thesis, Department of Automatic Control and Systems Engineering, The University of Sheffield, Sheffield, UK, 2003. [15] A. Wang, Hybrid modelling and decision support for ventilator management in intensive care units, Ph.D. Thesis, Department of Automatic Control and Systems Engineering, The University of Sheffield, Sheffield, UK, 2008. [16] A. Wang, G. Panoutsos, M. Mahfouf, G.H. Mills, An improved blood gas intelligent hybrid model for mechanically ventilated patients in the Intensive Care Unit, in: IASTED BioMED’07, Proceedings of the 5th IASTED International Conference on Biomedical Engineering, February 14–16, 2007, Innsbruck, Austria, 2007, pp. 73–78. [17] A. Wang, G. Panoutsos, M. Mahfouf, M. Mills, Goal-directed therapy for general ICU patients using aggregated multi-objective optimisation, in: 17th IFAC World Congress, 6–11 July, 2008, Seoul, Korea, 2008. [18] ARDSnetwork, Ventilation with lower tidal volumes as compared with traditional tidal volumes for acute lung injury and the acute respiratory distress syndrome, The Acute Respiratory Distress Syndrome Network, New England Journal of Medicine 342 (18) (2000) 1301– 1308. [19] D.E. Goldberg, Genetic Algorithms in Search, in: Optimization and Machine Learning, Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA, 1989.