Automation of a portable extracorporeal circulatory support system with adaptive fuzzy controllers.

G Model JJBE-2477; No. of Pages 10

ARTICLE IN PRESS Medical Engineering & Physics xxx (2014) xxx–xxx

Contents lists available at ScienceDirect

Medical Engineering & Physics journal homepage: www.elsevier.com/locate/medengphy

Automation of a portable extracorporeal circulatory support system with adaptive fuzzy controllers A. Mendoza García a,b,∗,1 , M. Krane b,1 , B. Baumgartner a , N. Sprunk a , U. Schreiber b , S. Eichhorn b , R. Lange b , A. Knoll a a

Faculty of Informatics, Department of Robotics and Embedded Systems, Technische Universität München, Munich, Germany German Heart Center Munich, Department of Cardiovascular Surgery, Department of Experimental Surgery, Technische Universität München, Munich, Germany b

a r t i c l e

i n f o

Article history: Received 12 May 2013 Received in revised form 16 April 2014 Accepted 26 April 2014 Keywords: Fuzzy control Control Extracorporeal circulation Cardiovascular modelling Adaptive control Extracorporeal support system Haemodynamic simulation Gas exchange modelling

a b s t r a c t The presented work relates to the procedure followed for the automation of a portable extracorporeal circulatory support system. Such a device may help increase the chances of survival after suffering from cardiogenic shock outside the hospital, additionally a controller can provide of optimal organ perfusion, while reducing the workload of the operator. Animal experiments were carried out for the acquisition of haemodynamic behaviour of the body under extracorporeal circulation. A mathematical model was constructed based on the experimental data, including a cardiovascular model, gas exchange and the administration of medication. As the base of the controller fuzzy logic was used allowing the easy integration of knowledge from trained perfusionists, an adaptive mechanism was included to adapt to the patient’s individual response. Initial simulations show the effectiveness of the controller and the improvements of perfusion after adaptation. © 2014 IPEM. Published by Elsevier Ltd. All rights reserved.

1. Introduction Cardiogenic shock caused by myocardial infarction is associated with a mortality of more than 60% [1]. Early treatment of patients having low cardiac output or during prolonged cardiopulmonary resuscitation (CPR) could be prevented from multi-organ failure with the use of an extracorporeal support system (ECSS) reducing this high mortality rate [2–4]. Since its first successful use in 1953, cardiopulmonary bypass (CPB) has been established as the gold standard for maintaining circulatory and pulmonary function in patients undergoing cardiac surgery [5,6]. Throughout the past decade, CPB has been introduced into cases of emergency circulatory resuscitation in patients with cardiogenic shock. The use of CPB systems in emergency situations is often limited by the large size and complicated setup of currently available

∗ Corresponding author at: Faculty of Informatics, Department of Robotics and Embedded Systems, Technische Universität München, Munich, Germany. Tel.: +49 89 289 18192. E-mail address: [email protected] (A. Mendoza García). 1 These authors contributed equally to this work.

systems [3]. Recently a variety of portable devices have become available for the treatment of patients suffering from cardiogenic shock outside the cardiac surgery unit. These devices are operated by trained medical staff rather than by perfusionists. To reduce the workload of the operator in a hectic scenario, and allow the safe transportation of patients to the hospital a complete automated perfusion system guided by an adaptive and robust control system is desirable. This paper describes the development of such control system for the automation of Lifebridge B2T, an already available portable extracorporeal circulatory support system (ECSS) [7,8]. Previous attempts to automate an extracorporeal support system have been performed by Misgeld et al. [9] and Meyrowitz [10]. Their studies focused on the automation of normal heart–lung machines (HLM) more commonly used in the operating room. These types of machines use rotary pumps and are not suitable for transportation. Misgeld proposed a proportional–integral (PI) and a robust type of controller and Meyrowitz presented a model predictive controller. This paper focuses on the automation of the more portable ECSS, which compared to the conventional HLMs uses a centrifugal pump to generate blood flow and is connected through femoral cannulation. The proposed automation is based on fuzzy logic, allowing the easy integration of the expert’s knowledge with the creation of

http://dx.doi.org/10.1016/j.medengphy.2014.04.009 1350-4533/© 2014 IPEM. Published by Elsevier Ltd. All rights reserved.

Please cite this article in press as: Mendoza García A, et al. Automation of a portable extracorporeal circulatory support system with adaptive fuzzy controllers. Med Eng Phys (2014), http://dx.doi.org/10.1016/j.medengphy.2014.04.009


ARTICLE IN PRESS A. Mendoza García et al. / Medical Engineering & Physics xxx (2014) xxx–xxx

2

control rules. Previous research has focused on creating a multiple input–single output adaptive mechanism that can change the controller’s rules based on a reference model and the response of the system. This was shown to work effectively for adjusting the pressure and flow of an in vitro system consisting of a pump and a reservoir acting as a compliance element [11]. This control mechanism was extended for this work to create a multiple input–multiple output automation system capable of adjusting flow, pressure, oxygen saturation and medication. For the development of the controller an animal model was used to obtain haemodynamic characteristics similar to patients under extracorporeal circulation (ECC). A mathematical model was constructed from the animal model and was used for extensive evaluations and validation of the controller. 2. Methods 2.1. Animal model Domestic pigs were used in the animal model for the acquisition of haemodynamic parameters during ECC. The experiments were approved by the Bavarian authorities, and the animals received humane care in compliance with Guide for the Care and Use of Laboratory Animals (NIH publication 85-23). Four domestic pigs weighing 50 ± 0.7 kg were pre-medicated with an intramuscular injection of ketamine (15, Ketanest® , Parke Davis, Munich, Germany) and an atropine sulphate injection (0.5,

Braun, Melsungen, Germany). General anaesthesia was induced by intravenous injection of propofol (60–100 mg, propofol, Lipuro, B.Braun Ag, Melsungen, Germany). Anaesthesia was maintained by continuous intravenous application of propofol (10 mg/kg/h propofol 2%) and fentanyl (30 ␮g/kg/h, fentanyl, Jannsen Cilag, Neuss, Germany) through a syringe pump. After endotracheal intubation, the pigs were placed on a respirator and ventilated with a mixture of oxygen and air. The fraction of inspired oxygen (FiO2 ) was set on 0.5. A catheter was inserted into the jugular vein (ArrowHowesTM Quad-Lumen central venous catheter, Arrow International Inc., USA) for monitoring of the central venous pressure (CVP). Through the right femoral artery, a catheter tip manometer (Millar MIKROTIP® SPC350, Houston, TX, USA) was placed in the descending aorta for monitoring the aortic pressure. Median sternotomy was done and the pericardium was opened. To measure the aortic flow, a perivascular ultrasonic flow probe (A-Serie, Transonic Systems Inc., Ithaca, NY, USA) was placed at the descending aorta above the crossing of the pulmonary veins. Another flow probe (C-Serie, Transonic Systems Inc., Ithaca, NY, USA) was placed in the ascending aorta. Connection with the ECSS was done through femoral cannulation. From the arterial side a 20 F arterial cannula (Medtronic, Inc., Minneapolis, MN, USA) was introduced into the femoral artery and a 22 F cannula (Edwards Lifesciences, CA, USA) was placed in the femoral vein (Fig. 1). For the acquisition of gas exchange information a CDI 500 gas analyser was used with a sampling rate of 1 sample every 6 s (Terumo Medical Corp, Tokyo, Japan). The sensors were placed

Fig. 1. Experimental setup and procedure.




between the arterial and venous cannulas and the ECSS. An oximeter (ChipOx, Corscience, Erlangen, Germany) was used to measure oxygen saturation at 1 Hz. Heart beat and ECG readings were obtained with a 4 lead ECG reader sampling at 200 Hz (EMI12, Corscience, Erlangen, Germany) was used placing gel electrodes. From the ECSS readings of pump flow, input, pre-oxygenator and output pressures were obtained. All of the sensor data was recorded using our developed software [12] (AutoMedic) and a 16bit data acquisition board set to acquire analogue data at 200 Hz (NI PCMCIA 6036E, National Instruments, TX, USA). The ECSS machine was previously primed with a saline solution and the centrifugal pump was set in a continuous closed circulation at 1000 rpm. After all the sensors were in place steady state data was acquired for 20 min (with arterial and venous cannulas clamped). The ECSS was then connected to the femoral cannulas and air was extracted from inside the tubing. The clamps were then slowly removed and extracorporeal circulation was started increasing the centrifugal pump speed until reaching an extracorporeal flow rate (EFR) of 4.5 l/m. At this point gas analysis could be started with the CDI 500. An additional 20 min of steady state with extracorporeal circulation was recorded at this stage. Afterwards the pump speed was changed in steps of 500 rpm starting from 2000 rpm to 3900 rpm with 5 min of recording in between. After the steady stage several operational tests were made to evaluate the interaction of the ECSS with the beating heart.

3

Gas concentrations were analyzed at all times and appropriate adjustments to the gas blender connected to the oxygenator were done by changing the FiO2 and gas flow to keep the arterial oxygen partial pressure at 150 mmHg and oxygen saturation at approx. 75% from the venous return. The CO2 partial pressures were kept at approx. 45 mmHg. For the evaluation of reactions to medication a vasoconstrictor, norepinephrine (NEP) (0.1 mg bolus; Arterenol, Sanovi Aventis, Frankfurt, Germany) was administered and afterwards a vasodilator, sodium nitroprusside (SNP) (10 mg bolus; Nipruss, UCB Pharma, Monheim, Germany). Fig. 2 shows the mean values of heart rate, arterial pressure and flow obtained from the different experiments. A comparison is shown before ECC and on ECC. The heart rate had a slight increase when ECC was started however there was no significant difference. The MAP in all experiments decreased even after having the ECSS blood pump at full speed starting at 75 ± 23 mmHg and being reduced to 61 ± 17 mmHg. From the arterial flow the ascending aorta started at 5.18 ± 1.2 l/min and was reduced to 2.1 ± 0.9 l/min. From the descending aorta an initial flow of 2.4 ± 1.4 l/min was registered and this decreased to −0.3 ± 0.3 l/min on ECC. With the heart stopped the average MAP decreased to 50 ± 10 mmHg. The flow at the ascending aorta showed a small negative flow of 0.3 ± 0.3 l/min which could be a possible indication of backflow to the direction of the heart. The descending aorta showed a negative flow of −1 ± 0.5 l/m indicating a flow going into the direction of the upper body.

Fig. 2. Experimental values before and during ECC.




4

Table 1 Centrifugal pump constants.

2.2. Mathematical model The mathematical model was composed of the cardiovascular system connected to the ECSS; this model calculates the haemodynamics and blood distribution. A gas exchange model consisting of six compartments was used for the calculation of oxygen delivery and extraction. A baroreceptor model was added to regulate the heart rate, and a medication model for possible vasoactive drugs delivery (Fig. 3). The cardiovascular system was characterized as a series of hydraulic compliant elements (C) interconnected by resistance elements (R). Additional inductance elements (L) were introduced between compartments to simulate the inertia of blood. The heart was composed of a four-compartment element with the corresponding valves represented as diodes. The relations of pressure (P), volume (V) and flow (Q) were introduced in relation to the neighbour components. These relations were expressed in the following formulas: Pi = (Vi − Vi−1 ) · Qi =

1 Ci

Pi − Pi−1 Ri

(1) (2)

dVi = Qi−1 − Qi dt

(3)

1 dQi = (Pi−1 − Pi ) · Li dt

(4)

The ventricles of the heart were modelled as a three-walled system, the right ventricular (RV) and left ventricular (LV) free walls and the coupling septal wall. This creates four functional volumes. The heart contractions were modelled by changing the contractility of each compartment and a sinusoid activation signal y(t) with a frequency provided by the baroreceptors to established the heart rate [13]. E(t)I = (EiMax − EiMin ) ∗ y(t)i + EiMin

(5)

The baroreceptor model takes as input the mean arterial pressure and the output is the instantaneous firing frequency of the

ki

Value

1 2 3 4

−7.85 × 10−5 0.0219 0.002 −2.00

baroreceptor. This is represented by a sigmoidal baroreflex curve (bfc) function as presented in Eq. (6), described by Ursino [14] and Brawany [15]. bfc(t) =

e(MAP − MAP0 ) 1 + e(MAP − MAP0 )

(6)

MAP0 is the nominal value of MAP; is an empirical constant (0.06263 mmHg−1 ). The pulmonary circulation consists of a single compliance and resistance. The systemic circulation was separated in two sections corresponding to the upper body, including the cerebral circulation and arms, and lower body for the rest. For each of these systemic sections three compartment elements were used to represent the arteries, capillaries and veins. From the lower systemic circulation before the arterial compartment and after the venous compartment the ECSS was connected with additional resistances representing the femoral vein and artery. From these two points the cannulas were attached. For the representation of the components of the ECSS a compliance was used to represent the reservoir and oxygenator. The centrifugal pump was modelled as a pressure generator element with a change of pressure between the inlet and the outlet depending on the speed at which it is running (ω) and the flow being produced (Q). Ppump = [Q · (k1 · ω + k2 ) + k3 · ω + k4 ] · · g

(7)

P is the pressure drop between the inlet and the outlet of the pump (mmHg), is the density of the blood and g is the gravity. The constants kn were obtained experimentally by generating different pump speeds and changing the resistance between the inlet and outlet, these values are shown in Table 1.

Fig. 3. Mathematical model. Subsystem diagram.


G Model

ARTICLE IN PRESS

JJBE-2477; No. of Pages 10

A. Mendoza García et al. / Medical Engineering & Physics xxx (2014) xxx–xxx

5

Table 2 ECSS component constants. i

Name

Ka

Kb

1 2 3 4 5 6

Reservoir Oxygenator Filter Tubing Venous cannula Arterial cannula

0.45 0.9 0.7 0 2.5 3.8

1.3 10 2.7 25 10 6

The components from the ECSS were analyzed separately to extract the pressure difference between the input and output of each component (P) in relation to flow (Q). This was achieved by connecting each component in line with the centrifugal pump. Pressure transducers were placed at the inlet and outlet together with an ultrasonic flow probe from Transonic. A quadratic function was used to model each element: P = Kai · Q 2 + Kbi · Q

(8)

The constants Ka and Kb were extracted for each component obtaining the values as shown in Table 2. For the gas exchange calculation (oxygen and carbon dioxide) a six compartment model was used for each gas: a systemic artery, capillary and vein compartments, a pulmonary compartment, a reservoir and a compartment for the oxygenator in the ECSS. The change of partial pressure of each gas was calculated based on the gas concentration in every compartment with the following formulas: q (c − ci ) dCi = i i−1 Vi dt

(9)

Ci corresponds to the concentration of the corresponding gas (O2 or CO2 ) in compartment i. q is the blood flow going into the compartment and Vi is the total amount of blood of compartment i. The relation between oxygen concentration and partial pressure is calculated based on the haemoglobin concentration and oxygen saturation: CtO2,i = Hb × 1.34 × SpO2 + PO2,i × 0.003

(10)

CtO2 is measured in mlO2 /dl, Hb is the haemoglobin concentration measured in g/dl; 1.34 is the oxygen binding capacity of blood with units mlO2 /gmHb, SpO2 is the oxygen saturation, PO2 is the oxygen partial pressure in mmHg and 0.003 is the solubility of oxygen in the blood in mlO2 /mmHg/dl. The gas exchange in the oxygenator was modelled considering the partial pressures of the blood compartment with an additional gas compartment that is connected to an air-oxygen blender [10]. dPG,i V˙ G · (PG,i−1 − PG,i ) + Patm · DG (BlPG,i−1 − PG,i ) = VG,i dt

(11)

The sub index G corresponds to either O2 or CO2 . VG is the flow of the gas mixture introduced into the oxygenator and is determined by the oxygen blender. PG is the input partial pressure of O2 (PO2 ) and CO2 (PCO2 ). DG is the diffusion factor for O2 and CO2 . BlPG corresponds to the gas partial pressure of the blood entering the oxygenator. The input partial pressure PCO2 is 0 mmHg. The input oxygen partial pressure is based on the configuration of the FiO2 in the blender and is calculated with the following formula: PO2,in = FiO2 · PO2,ref

(12)

PO2,ref corresponds to the atmospheric pressure (760 mmHg). The vasoactive drugs were modelled using a similar configuration of six compartments used in the gas exchange model, with the calculation of blood flow obtained from the cardiovascular model. The volume loss was calculated based on the half-life of each drug. Sodium nitroprusside (SNP) was used as a vasodilator and as

Fig. 4. Upper: sensor location; lower: pressure and flow curves related to pump speed, obtained experimentally (dots) and through simulation (lines).

vasoconstrictor norepinephrine (NEP). The effects of the vasoconstrictor and vasodilator were calculated with the following formulas: Ra = R0 + Cta,d · Krd

(13)

Ca = C0 + Cta,d · Kcd

(14)

Ra and Ca are the resistance and compliance value of the arterial vein respectively. R0 and C0 are the base values, Cta is the drug concentration in the arterial compartment and Kr and Kc are constants for each drug. The parameters of the mathematical were adjusted based on the data obtained from the animal experiments. This was done by changing the resistances in the model together with the compliances. The contractility of the heart was also adjusted to resemble the flows and pressures that were obtained experimentally. The simulation of haemodynamic parameters compared to experimental data is shown in Fig. 4 together with the location of the sensors in the ECSS. The top graph in Fig. 4 shows the pressures in the different parts of the ECSS, the middle graph shows the mean arterial pressure (MAP) and the last one shows the different flow rates: extracorporeal flow rate (EFR) produced by the machine, flow at the ascending aorta (FAA) and flow at the descending aorta (FAD). The horizontal axis represents the centrifugal pump speed in rpm. The flow




6

Fig. 5. Simulation results compared to experimental data.

relations of the ascending and descending aorta show at maximum speed a reduction of the cardiac output from 3 l/min to 1.5 l/min. The flow of the descending aorta is negative, as described in Fig. 2. Fig. 5a shows the arterial pressure from the experimental data (continuous) compared with the arterial pressure obtained through simulation (dotted). Fig. 5b shows the flow generated by the centrifugal pump and the pressure between the inlet and outlet at different speeds. Fig. 5c shows the effect of arterial pressure after a dosage of a vasoconstrictive substance was introduced and Fig. 5d shows the effect of a vasodilator. Shown in grey is the continuous curve obtained experimentally and the black dotted line shows the mean pressure of the simulation model.

2.3. Adaptive Fuzzy Control The basic components of a fuzzy controller are shown in Fig. 6a. An adaptive mechanism improves the response of the controller based on the response of the patient. This was achieved following the approach of a model reference control [16]. The adaptive fuzzy controller (AFC) architecture is shown in Fig. 6b, consisting of three main components. The first is a set of reference models used for each input parameter. The second is a sub set of fuzzy controllers that contain the rules dictated by the perfusionist. Each sub controller defines how a control variable should be changed according to specific input signals. These sets of controllers are defined as knowledge controllers. The third component is an adaptive controller. This controller starts with an empty rule set. It is composed of the same inputs as the knowledge controllers plus additional signals that may be relevant to the parameter to control. The knowledge controllers may be composed of predefined rule structures that define the behaviour of the controller, for example a fuzzy-PI type of controller may be used for one input, the rule structure for this type of controller is further described by Li [17].

Control parameters are defined as vital signals that are maintained at a given range to assure organ perfusion. The values of the different parameters are predefined by the operator and are referred to as target values. A target error (tE) is defined as the difference between the target value and the current value of the patient. A second order function was used as a reference model to indicate the controller how the current signals should reach its predefined target. The difference between the reference model and the current signal value is called reference error (rE). These errors are used as inputs of the different controllers (Fig. 6c). The input and output sets were chosen of triangular shape to reduce computational cost. The sets are spread covering the range of values configured for each input. In the knowledge controllers 5 triangular sets are used for each input and 11 sets for the outputs. In the adaptive controllers 7 sets are used for the inputs and 31 triangular sets are used for the output to enable the adaptive algorithm to do small corrections. All of the fuzzy controllers are of type Mamdani using Min–Max inference method and Centroid for defuzzification. During the automation process, when the current input values are introduced into the adaptive controller and there exists no rule exceeding a predefined threshold a learning process takes place where the knowledge controllers are activated and a new rule is created. If a rule already exists an adaptive process is executed to modify the rules and improve control performance. The output obtained from the adaptive controller is then used to increase or decrease the value of the control variables. Further details of the learning and adaptive process are explained in more detail in Ref. [11]. The automation of the ECSS was accomplished by using four adaptive controllers: one in charge of adjusting the centrifugal pump speed, a second for the gas blender FiO2 , the third for the gas flow, and the last one for the administration of medication. Table 3 describes the selected input parameters with a configured


G Model

ARTICLE IN PRESS



7

Fig. 6. (a) Fuzzy control; (b) control signals; (c) adaptive mechanism.

Table 3 Input and control variables. Name

Min

Normal

Max

Units

80 6 1600 100 80 200 60 50 50

mmHg l/min dyn s/cm5 % % mmHg mmHg mmHg mmHg

3900 100 10 2 0.1

rpm % l/min ␮g/kg/min ␮g/kg/min

Input parameters MAP EFR SVR SpO2 a SpO2 v PO2 a PO2 v PCO2 a PCO2 v

50 3 500 70 50 90 30 30 30

60 5 960 99 80 100 40 40 41

Control variables Pump speed FiO2 Gas vol. SNP NEP

1000 20 2 0 0

– – – – –

minimum, maximum and target value. These values may change depending on the precondition and medical history of the patient. Fig. 7 shows the reference models used in the automation system, with the four adaptive controllers created, described as follows. The main inputs of the centrifugal pump controller are MAP and EFR. The control rules consist of increasing pump speed if MAP or EFR are lower than the target value and decrease pump speed if MAP or EFR are higher. Oxygen saturation is also used as input were the

Fig. 7. Reference models and controllers used for automation of ECSS.

pump speed may be increased to generate more flow if the FiO2 is already at 100% with a gas flow at 10 l/min and the venous oxygen saturation is lower than the target value. Fig. 8 shows the pump control structure. The gas blender controller monitors the oxygen and carbon dioxide in the arterial and venous side provided by the gas analyser. The controller is separated in a set of rules for the FiO2 and another for the gas flow. The FiO2 rules are used to control oxygen concentrations: If the oxygen saturation is low in the venous side the FiO2 is increased, if the oxygen saturation is normal in the venous side and the oxygen partial pressure is high on the arterial side then the FiO2 is reduced. The gas flow rules consider the carbon dioxide concentration: If the PCO2 in the venous side is higher than the pre-set target then the flow is increase, on the contrary if it is lower, then it is slowly decreased. Additional rules increase the gas flow if the FiO2 is at 100% and the venous SpO2 is low. For the medication controllers a systemic vascular resistance (SVR) was approximated by considering the MAP together with the CVP and replacing the cardiac output with the EFR:

SVR ≈

MAP − CVP EFR

(17)

Once the pump speed is stabilized the medication control is activated. A vasodilator and vasoconstrictor is used to produce contrary effects and are exclusive: If one substance is used the other substance is not. A unified medication variable was used (Meds). When this variable is positive this corresponds to the vasoconstrictor dosage, while the vasodilator remains on 0; if the variable is negative it corresponds to the vasodilator using an absolute value. The rules of the controllers consist on modifying this variable: if the current SVR is higher than the target SVR then Meds is decreased, if the current SVR is lower than the target then Meds is increased. For the implementation of the mathematical model a Mathematical Modelling Language (MML) was used to describe the model which was compiled and executed by the simulation system called Jsim. This was extended with a simulation engine capable of generating a continuous simulation while providing an interface to the control system. The controller was programmed in C++ and was integrated into the software library AutoMedic previously used to capture the experimental data [18].


G Model

ARTICLE IN PRESS



8

Fig. 8. Control structure for automation of pump speed based on pressure, flow and oxygen saturation.

Gas Exchange

Target FC / no Meds FC / Meds

100

Reference AFC / no Meds AFC / Meds

SpO2v [mmHg]

MAP [mmHg]

Hemodynamics

80 60 40

Target FC / no Meds FC / Meds

100

Reference AFC / no Meds AFC / Meds

80 60 40 20 60

6 PCO2v

EFR [l/min]

8

4 2

40

0

100 FiO2 [%]

SVR

1500 1000 500

80 60 40 20

GasFlow [L/min]

PumpSpeed [rpm]

10

4000

3000

8 6 4 2 0

2000

0

5

10

15 time [min]

20

25

30

ISDN [ug]

30

20

10

0 0

5

10

15

20

25

30

time [min]

Fig. 9. Simulation results with controller.


G Model

ARTICLE IN PRESS



4. Discussion

2000 AFC / no Meds AFC / Meds

accMSE

FC / no Meds FC / Meds

0 0

5

10

15

20

25

30

time [min]

Fig. 10. Accumulative mean square error for the comparison of different controllers.

3. Results For the evaluation of the proposed controller four simulations were carried out. The first (FC/noMeds) consists of only using the knowledge fuzzy controllers without adaptation and without medication control. The second consists of the activation of the adaptive controllers also without medication (AFC/noMeds). The third uses the fuzzy controllers with the medication control activated (FC/Meds) and the last shows the adaptive fuzzy controllers and medication (AFC/Meds). From the simulation model the contractility of the heart were reduced to produce low cardiac output and the resistance of the systemic artery was increased to produce a high pressure. The controllers in all of the simulations were configured with the target values shown in Table 3. The adaptive controllers were previously trained with several simulations to learn and adapt to the behaviour of the system. These results are shown in Fig. 9. The simulation shows that when the medication controllers were not activated, due to a high SVR, the values of MAP and EFR could not be achieved. An equilibrium point is reached where the MAP is slightly higher (+5 mmHg) than its target and the EFR is lower (−1 l/m). This was corrected with the activation of the medication control by decreasing the SVR and increasing the pump speed to reach the target MAP and EFR. In the case of control without adaptation the target value of SpO2 v was exceeded by 10% and oscillations around the target value were generated. The adaptive control was able to reduce these oscillations and subsequently achieve the target values. Once the SpO2 v was achieved by changing the FiO2 the gas flow was slowly increased to reduce the PCO2 . An accumulative mean square error (accMSE) was calculated to compare the performance of each controller, consisting on calculating the error between the current values and the target values. This error is accumulated over time (Fig. 10): MSE(parami ) = (ti − currenti )2 accMSE(t) = ˙[MSE(parami )] + accMSE(t − 1)

9

(18)

This analysis shows that the adaptive fuzzy control with medication (AFC/Meds) had the lowest accMSE. FC/Meds had the highest error since the response of the medication caused the pressure to drop and the EFR to increase, generating oscillations without reaching stability. The controllers without medication had a constant increase of MSE since the target values of EFR and MAP cannot be achieved due to a high SVR. The adaptive controllers however gave a better response.

The evolution and miniaturization of the ECSS has allowed these systems to be taken directly to patients suffering from cardiogenic shock, reducing the time to treatment which can be crucial to improve the chances of survival [19,20]. This however requires the system to be easy to use for a non-perfusionist and to have been rigorously tested for safety and efficacy. The automation of such a system may help in giving optimal perfusion to the patient while allowing paramedics to focus on the patient during transportation. The amount of EFR that the machine produces will change from one patient to another. This variations will depend if the heart is still beating or not, the heart strength when beating and the vascular resistance. In the case of the beating heart, in the interaction between the EFR and the cardiac output it was noticed that the load of the heart was effectively reduced. From the different experiments it was noticed that it was not possible to achieve EFR greater than 5 l/min and MAP greater than 80 mmHg even if the centrifugal pump was running at full speed. Further analysis indicated that the resistance produced by the cannulation caused a considerable pressure drop; several studies have already focused on reducing cannula resistance [21]. Several cardiovascular models have previously been developed [22–25], in these models there is a trade-off between detail of the design model, increasing complexity and the computational requirements, or if the system is too simple it will not be able to fully represent the desired behaviour. The presented model is capable of recreating the experimental data while keeping low model complexity. With the division of the systemic circulation into upper and lower sections it was possible to analyze the flow interaction between the beating heart and the centrifugal pump. Regarding the controller, several in vitro studies done in our research group [26] show that a classic PI controller may have a better performance in the control of one specific parameter compared to a fuzzy controller, once the PI controller gains have been properly adjusted, however when the number of variables considered for control are increased and the model parameters are modified the fuzzy controller provides a more feasible solution. The presented simulation scenario and additional simulations have shown that the adaptive fuzzy controller was capable of reaching the desired values providing optimal perfusion to the organs. A proper stability analysis is required for the presented controller. Several proposed methods have been presented by different authors [27,28] for simple fuzzy systems. We have expressed the potential benefits of automating an ECSS and the steps that were followed in the design of a controller. The animal experiments provided the reference data necessary to create the mathematical model and understand the haemodynamic behaviour of the cardiovascular system together with the ECSS. The simulations done with the mathematical model were capable of representing the experimental data and allowed intensive evaluation and optimization of the developed controller. Fuzzy logic allowed a straight forward translation of the decisions made by perfusionists into control rules, and the adaptive mechanism was capable of improving the controller behaviour by making proper adjustments to the rule base. Further tests and validation are required to make the controller available to patients, including safety mechanisms which will be the focus of further work.

Acknowledgement JSim is provided free for non-commercial use as a public service by the National Simulation Resource at the University of Washington (http://nsr.bioeng.washington.edu/jsim/).


G Model JJBE-2477; No. of Pages 10 10


Funding This work has been supported by an unrestricted educational grant from Bayerische Forschungsstiftung (AZ-699-06). Ethical approval Not required. Conflict of interest None declared. References [1] Carnendran L, Abboud R, Sleeper LA. Trends in cardiogenic shock: report from the SHOCK study. Eur Heart J 2001;22:472–8. [2] Chen YS, Lin JW, Yu HY. Cardiopulmonary resuscitation with assisted extracorporeal life-support versus conventional cardiopulmonary resuscitation in adults with in-hospital cardiac arrest: an observational study and propensity analysis. Lancet 2008;372:554–61. [3] Chen YS, Chao A, Yu HY. Analysis and results of prolonged resuscitation in cardiac arrest patients rescued by extracorporeal membrane oxygenation. J Am Coll Cardiol 2003;41:197–203. [4] Maif P, Hoermann C, Moertl M, Bonatti J, Falbesoner C, Balogh D. Percutaneous venoarterial extracorporeal membrane oxygenation for emergency mechanical circulatory support. Resuscitation 1996;33:29–34. [5] Krane M. Die Herz-Lungen-Maschine. In: Wintermantel E, editor. Medizintechnik life science engineering. Berlin, Heidelberg: Springer Verlag; 2008. p. 1069–82. [6] Gravlee GP, editor. Cardiopulmonary bypass: principles and practice. Philadelphia: Lippincott Williams & Wilkins; 2007. [7] Krane M, Mazzitelli D, Schreiber U, Mendoza Garcia A, Braun S, Voss B, et al. Lifebridge B2T a new portable cardiopulmonary bypass system. ASAIO J 2010;56:52–6. [8] Krane M, Mazzitelli D, Schreiber U, Mendoza Garcia A, Voss B, Badiu C, et al. First experience with a new portable cardiopulmonary bypass system – lifebridge B2T with percutaneous femoral cannulation. Comput Cardiol 2008:269–72. [9] Misgeld B, Werner J, Hexamer M. Robust and self-tuning blood flow control during extracorporeal circulation in the presence of system parameter uncertainties. Med Biol Eng Comput 2005;43:589–98. [10] Meyrowitz G. Automatisierung der Herz-Lungen-Maschine. Mensch-undBuch-Verlag; 2005. [11] Mendoza Garcia A, Sprunk N, Baumgartner B, Schreiber U, Eichhorn S, Krane M, et al. Application of adaptive fuzzy controllers for the automation of medical devices. In: FUZZ-IEEE. 2012. p. 1–5. [12] Mendoza Garcia A, Baumgartner B, Schreiber U, Krane M, Knoll A, Bauernschmitt R. Automedic: fuzzy control development platform for a mobile heart–lung machine. IFMBE Proc 2009;25(7):685–8.

[13] Rideout VC. Mathematical computer modeling of physiological systems. Englewood Cliffs, NJ: Prentice Hall; 1991. p. 261 pp. [14] Ursino M. Interaction between carotid baroregulation and the pulsating heart: a mathematical model. Am J Physiol Heart Circ Physiol 1998;275(5): 1733–47. [15] Brawany MA. A functional cardiovascular model with IV cardiac drugs action. Electron Electr Eng 2009;1(1):35–42. [16] Passino KM, Yurkovich S. Fuzzy control. Addison-Wesley; 1998. [17] Li H-X. Approximate model reference adaptive mechanism for nominal gain design of fuzzy control system. IEEE Trans Syst Man Cybern B Cybern 1999;29(1):41–6. [18] Mendoza Garcia A, Baumgartner B, Schreiber U, Krane M, Bauernschmitt R, Knoll A. Simulation of extracorporeal circulation for the design of a fuzzy controlled perfusion. IASTED Biomed Eng 2010;1:2. [19] Lee S, Hong Y. Extracorporeal life-support in patients requiring CPR. Lancet 2008;372. [20] Shin JS, Lee S-W, Han G-S, Jo W-M, Choi S-H, Hong Y-S. Successful extracorporeal life support in cardiac arrest with recurrent ventricular fibrillation unresponsive to standard cardiopulmonary resuscitation. Resuscitation 2007;73(2):309–13. [21] von Segesser LK, Kalejs M, Ferrari E, Bommeli S, Maunz O, Horisberger J, et al. Superior flow for bridge to life with self-expanding venous cannulas. Eur J Cardiothorac Surg 2009;36:665–9. [22] Olansen J, Clark J, Khoury D, Ghorbel F, Bidani A. A closed-loop model of the canine cardiovascular system that includes ventricular interaction. Am J Physiol Heart Circ Physiol 2001;281:2661–79. [23] Bauernschmitt R, Naujokat E, Mehmanesh H, Schulz S, Vahl C, Hagl S, et al. Mathematical modeling of extracorporeal circulation: simulation of different perfusion regimens perfusion. Perfusion 1999;14: 321–30. [24] Clark LUK, Ghorbel JW, Ware FH, Zwischenberger DL, Bidani JB. A whole-body gas exchange in human predicted by a cardiopulmonary model. Cardiovasc Eng 2003;3(1):1–19. [25] Luo C, Ware D, Zwischenberger J, Clark Jr JW. Using a human cardiopulmonary model to study and predict normal and diseased ventricular mechanics, septal interaction, and atrio-ventricular blood flow patterns. Cardiovasc Eng 2007;7(17):31. [26] Baumgartner B, Mendoza A, Eichhorn S, Schreiber U, Knoll A. A comparative study on extra-corporal circulation control. In: Conf. Proceedings IEEE EMBC. 2011. p. 4287–90. [27] Yi SY, Chung MJ. Stability analysis of a fuzzy logic controller for an uncertain dynamic system. In: Proceedings of the third IEEE conference on fuzzy systems, 1994 IEEE World Congress on computational intelligence, vol. 2. 1994. p. 1028–34. [28] Choi BJ, Kwak SW, Kim BK. Design and stability analysis of single-input fuzzy logic controller. IEEE Trans Syst Man Cybern B Cybern 2000;30(2): 303–9.


Performance and Safety of an Integrated Portable Extracorporeal Life Support System for Adults.

Autonomous navigation system using a fuzzy adaptive nonlinear H∞ filter.

Donation after circulatory determination of death: the university of michigan experience with extracorporeal support.

Adaptive Automation and Mental Processes.

Genetic optimization of fuzzy fractional PD+I controllers.

Extracorporeal circulatory management of priapism.

Type-II fuzzy decision support system for fertilizer.

Use of extracorporeal membrane oxygenation for mechanical circulatory support in a patient with 5-fluorouracil induced acute heart failure.

Universal fuzzy integral sliding-mode controllers for stochastic nonlinear systems.

Universal fuzzy models and universal fuzzy controllers for discrete-time nonlinear systems.

A fuzzy limbic system.

A portable TDCR system.

Percutaneous bi-atrial extracorporeal membrane oxygenation for acute circulatory support in advanced heart failure.

Evolving experience with mechanical circulatory support.

A contemporary review of mechanical circulatory support.

Acoustic analysis of a mechanical circulatory support.

Mechanical circulatory support.

Mechanical circulatory support.

Pediatric mechanical circulatory support.

Assessing experience in the deliberate practice of running using a fuzzy decision-support system.

Extracorporeal life support: experience with 2,000 patients.

Mechanical Circulatory Support in Children.

Mechanical circulatory support in pediatrics.

Adaptive neuro-fuzzy inference system for real-time monitoring of integrated-constructed wetlands.