Yao Sun MD, Isaac Kohane MD, PhD, Ann Stark MD
Children's Hospital, Harvard Medical School, Boston, MA
The control of oxygen delivery to mechanically ventilated newborn infants is a time intensive process that must balance adequate tissue oxygenation against possible toxic effects of oxygen exposure. Investigation in computer assisted control of mechanical ventilation is increasing, although there are very few studies involving newborn infants. We have implemented a fuzzy controller for the adjustment of inspired oxygen concentration (FIO2) in ventilated newborns. The controller utilizes rules produced by neonatologists, and operates in real-time. A clinical trial of this controller is currently taking place in the neonatal intensive care unit (NICU) of Children's Hospital, Boston, MA.
Oxygen toxicity plays a role in the development of chronic lung disease in newborn infants requiring mechanical ventilation. [1,2] In premature infants, varying levels of oxygen exposure are implicated in the development of retinopathy of prematurity. [3] Because of these effects, control of oxygen delivery to ventilated newborns has become a priority in neonatal intensive care.
Among the many ventilator parameters that affect patient respiratory status, the inspired oxygen concentration (FIO2) is most commonly adjusted on an acute basis to control oxygen delivery and maintain patient oxygen saturation levels. Manual control of the FIO2, however, may lag the clinical condition of the patient. That is, a patient may have an increased oxygen requirement as demonstrated by a lower oxygen saturation, but the manual increase of FIO2 may be delayed by human response times (i.e. a clinician may not be present to respond immediately). Conversely, a patient may have a decreased oxygen requirement as clinical conditions improve, yet the amount of oxygen delivered may not be immediately decreased. The latter scenario may be more common because of the perception that a patient with high oxygen saturation is "doing well" and does not require immediate intervention.
We have designed and implemented a microcomputer based system to automatically and continuously control the FIO2 delivered to mechanically ventilated newborn infants. This system utilizes a fuzzy logic controller based on "rules" generated by neonatologists who routinely provide care for ventilated infants. The goal of this control system is to maintain patient oxygenation (measured by oxygen saturation using pulse oximetry) at a target level set by the physician.
Instead of controlling the ventilator directly, the system currently operates by displaying suggested FIO2 changes to the physician, who then decides whether to execute the recommended change. This ensures medical safety until the system is fully tested for clinical efficacy. A clinical trial of the FIO2 control system is currently taking place in the neonatal intensive care unit (NICU) of Children's Hospital, Boston, MA.
Computer assisted ventilation
Investigation into computer-controlled or computer-assisted mechanical ventilation is expanding. One form of computer assistance is an "expert system" designed to advise the clinicians about ventilator management. Some recent examples include: VentPlan, a ventilator management advisor that interprets patient physiologic data to predict the effect of proposed ventilator changes [4]; ESTER, a program which assesses the patient's pathophysiologic state using modified APACHE-II criteria, then offers suggestions for weaning from intermittent mandatory ventilation [5]; WEANPRO, a program designed to help wean post-operative patients from ventilators [6]; and KUSIVAR, a program which describes a comprehensive system for respiratory management during all phases of pulmonary disease. [7]Although many such expert systems have been described, few have been tested in clinical patient care.
Other investigators have studied direct computer control of specialized aspects of ventilator management. For example, studies of computer-controlled optimization of positive end-expiratory pressure, and computerized protocols for management of adult respiratory distress syndrome have been explored by East. [8] A computerized ventilator weaning system for post-operative patients has been tested by Strickland. [9]
Experience in computer controlled ventilation in infants, however, is limited. In one of few reports available in the literature, Morozoff and Evans showed that their computerized FIO2 controller could maintain the oxygen saturation (SaO2) of a ventilated newborn infant for approximately 1 hour periods with results comparable to manual FIO2 control. [10]
Morozoff and Evans describe their FIO2 controller as a "differential-feedback" controller. Other investigators have described similar FIO2 controllers for adults based on the "proportional-integral-derivative" (PID) design. [11] For best response, most PID controllers and feedback-loop controllers need to have their control parameters optimized for the system in which they are used. This may lead to degradation of performance if the system changes (e.g. if the patient's physiologic status changes, or the controller is switched to a new patient). Yu addressed this problem in FIO2 control by using multiple controllers that dynamically adapted by selectively utilizing the controller that best matched the system response at any given point in time. [12]
Since Zadeh first published his seminal paper on fuzzy sets in 1965 [13], applications utilizing fuzzy logic have proliferated rapidly. Mamdani's development of fuzzy controllers in 1974 [14] gave rise to the utilization of these controllers in ever expanding capacities, particularly in Japan where many industrial processes now employ fuzzy control. [15] In addition, fuzzy control techniques have recently been applied to various medical processes, such as pain control [16]and blood pressure control. [17]
When compared to classical control theory, a fuzzy logic approach to control offers the following advantages: [15,18,19]
In the context of FIO2 control in the newborn infant, a fuzzy logic approach can simplify the many complex factors and interactions that determine patient oxygenation. For example, a ventilated infant may exhibit decreased oxygen levels in the blood (as measured by SaO2) for any of the following reasons: failure to make respiratory effort, a plug in the endotracheal tube, or an increase in pulmonary shunting. Each cause may require differing changes in FIO2 to maintain target SaO2 levels, and many other factors may influence oxygenation. At different times, the same magnitude of change in FIO2 may result in completely different oxygenation states, even within the same patient.
FIO2 control in the newborn thus demonstrates some of the previously mentioned features which make classical control techniques difficult to apply: the system to be controlled is complex with may factors and interactions, it is very difficult to model mathematically, and system responses to FIO2 changes are often non-linear and unpredictable.
It is noteworthy that our initial approach to the problem of FIO2 control involved the construction of a modified adaptive feedback-loop controller. After many weeks spent grappling with the previously defined problems and continually modifying the controller, clinical testing on ventilated infants revealed that our "classical" controller did not perform any better than standard manual control. In contrast, the fuzzy controller was designed and implemented within a week, and preliminary results of clinical testing show promise that the controller maintains target oxygenation saturations better than manual control while reducing oxygen exposure.
FIO2 Controller
We chose SaO2 as our measurement parameter and FIO2 as our control parameter for the operational model of maintaining patient oxygenation.
SaO2 as measured by pulse oximetry is a well established method of following patient oxygenation status. It's advantages over direct measurement of blood oxygen levels include rapid equilibrium with changes in blood oxygen levels, continuous monitoring, and it is noninvasive. We used the error between the patient's SaO2 and the target SaO2 ([[Delta]]SaO2), and the slope of SaO2 (SaO2-slope) as the specific inputs to the fuzzy controller.
Although many ventilator parameters affect patient oxygenation (e.g. mean airway pressure, ventilatory rate, tidal volumes, etc.), the FIO2 is used on maintain the desired oxygenation status when the patients overall respiratory status has been stabilized.
The design of the fuzzy controller then follows standard methods, with fuzzification of the input parameters, construction of fuzzy inference rules, and defuzzification or calculation of a "crisp" output value that represents the controller's action.
To fuzzify the input parameters, the values of [[Delta]]SaO2 and SaO2-slope were divided into fuzzy regions, with 7 regions chosen for [[Delta]]SaO2 and 5 regions chosen for the SaO2-slope. Triangular membership functions were assigned to each region, as illustrated in Figure 1.
Using the fuzzy input parameters, the inference rules that form the body of the controller were constructed in the standard declarative form: IF situation THEN action . The combination of 7 [[Delta]]SaO2 fuzzy regions and 5 SaO2-slope fuzzy regions yields 35 rules. The logic of these inference rules are based on the expert knowledge of the neonatologists. Some example rules follow:
Rule: IF the [[Delta]]SaO2 is small-negative
AND the SaO2-slope is medium-negative
(situation)
THEN increase the FIO2 by a
medium-positive amount.
(action)
Rule: IF the [[Delta]]SaO2 is large-negative
AND the SaO2-slope is large-negative
THEN increase the FIO2 by an
extremely-large-positive amount.
Rule: IF the [[Delta]]SaO2 is small-negative
AND the SaO2-slope is small-positive
THEN do nothing.
All 35 rules are summarized in Table 1.
For any pair of [[Delta]]SaO2 and SaO2-slope inputs, we apply each of the inference rules in turn. Each rule will yield an action value. The defuzzification step then involves choosing a method to combine all the action values into a final value (a "crisp" value) that represents the controller output. We used the weighted mean of all the rule outputs to produce a single output value, in this case a change in the FIO2. [20]
Although there are relatively few fuzzy inference rules, continuously calculating the crisp output in real-time may not always be feasible. To help minimize time-delays, we compiled the fuzzy inference rules into a look-up table at runtime. Thus, during actual fuzzy control operation, evaluating the inputs becomes a simple and fast table look-up producing the controller output.
The actual operation of the FIO2 controller is as follows:
1) SaO2 values are obtained for the patient every 1-2 seconds.
2) Every 10 seconds, the [[Delta]]SaO2 and the SaO2-slope are calculated.
[[Delta]]SaO2 = (ave. SaO2 values over last 10seconds)
- (target SaO2)
SaO2-slope = least squares regression of SaO2
values over last 10 seconds
3) The calculated [[Delta]]SaO2 and SaO2-slope are used as indices for the compiled fuzzy controller look-up table. A suggested FIO2 change is returned as the controller output.
System Components
The FIO2 fuzzy control system is implemented on an Apple Macintosh and is programmed in Macintosh Common Lisp. The SaO2 data is obtained from a Nellcor N-200 pulse oximeter through a RS-232 serial port on the back of the oximeter.
Study Design
In order to test the FIO2 control system in a medically safe manner, the computer was not allowed to directly control the ventilator oxygen delivery. Instead, suggestions for changes in FIO2 are displayed for the physician to execute according to his/her best medical judgment. The computer system was programmed to automatically record all recommended and actual FIO2 changes.
The clinical trial protocol was approved by the Clinical Investigations Committee of Children's Hospital, Boston, MA., and informed consent was obtained from the parents of patients entered into the study. Patients were eligible if they were newborn infants admitted to the NICU and required mechanical ventilation. Patients were excluded if they had demonstrated intracardiac shunting of blood from right to left, or if they required vasoactive pressor medications to maintain blood pressure.
Each infant was studied for a 6 hour period of time. The initial 2 hours served as a control period during which the computer system collected SaO2 and FIO2 data. No interventions were made during this time. For the subsequent 2 hour experimental period, the system made recommendations for FIO2 changes in addition to acquiring data. The investigator manually carried out the recommended FIO2 changes if they were consistent with his/her clinical judgment. Finally, another 2 hour control period of data gathering (without recommendations for FIO2 change)) completed the study period for the patient.
All clinical care activities proceeded as usual, and the NICU medical and nursing staff were not prevented from manually adjusting the FIO2 at any time during the trial.
Study Results
(Results to be inserted here)
Controlling oxygen exposure in newborn infants is a delicate balance. The infants must receive enough oxygen to ensure adequate tissue oxygenation and to prevent ischemia. Conversely, too much oxygen may produce toxic effects.
The FIO2 fuzzy controller shows promise in the preliminary trials to control patient oxygen saturation levels. It was able to maintain a target SaO2 better than routine manual control, and it reduced the overall oxygen exposure. Further clinical trials will test the actual clinical efficacy of this FIO2 controller, and additional patient data will allow more fine tuning of the fuzzy control parameters (e.g. the shape of the membership functions and the choice of fuzzy regions).
The ease of implementing this fuzzy controller illustrates some of the advantages of this approach. No complex mathematical models were required, the simple rule-based nature of the controller is easy to understand and modify, expert knowledge about the problem is utilized, and the controller was easily designed for non-linear system responses.
Current research in fuzzy control include combining it with other techniques such as neural networks and genetic algorithms [21,22], and adaptive or self-modifying fuzzy control. [20,23] As more medical processes become candidates for computerized control, the numerous options offered by these approaches will enhance the ability to produce a safe and clinically efficacious control system.
[1]Saugstad, O. Oxygen toxicity in the neonatal period. Acta Paediatr Scand1990, 79, 881.
[2]Wispe, J.; Roberts, R. Molecular basis of pulmonary oxygen toxicity.14 1987,
[3]Avery, G.; Glass, P. Retinopathy of prematurity: what causes it?Clin Perinatol 1988, 15,
[4]Rutledge, G.; Thomsen, G. Ventplan: a ventilator-management advisor.Proc Annu Symp Comput Appl Med Care 1991, 869.
[5]Hernandez-Sande, C.; Moret-Bonillo, V.; Alonso-Betanzos, A. ESTER: an expert system for management of respiratory weaning therapy.IEEE Trans Biomed Eng 1989, 36, 559.
[6]Tong, D. Weaning patients from mechanical ventilation. A knowledge-based system approach.Comput Methods Programs Biomed 1991, 35, 267.
[7]Rudowski, R.; Frostell, C.; Gill, H. A knowledge-based support system for mechanical ventilation of the lungs. The KUSIVAR concept and prototype.Comput Methods Programs Biomed 1989, 30, 59.
[8]East, T.; Bohm, S. A successful computerized protocol for clinical management of pressure control inverse ratio ventilation in ARDS patients.Chest 1992, 101, 697.
[9]Strickland, J.; Hasson, J. A computer-controlled ventilator weaning system.Chest 1991, 100, 1096.
[10]Morozoff, P.; Evans, R. Closed-loop control of SaO2 in the newborn infant.Biomed Instrum Technol 1992, 26, 117.
[11]O'hara, D.; Bogen, D.; Noordergraaf, A. The use of computers for controlling the delivery of anesthesia.Anesthesiology 1992, 77, 563.
[12]Yu, C.; He, W. Improvement in arterial oxygen control using multiple-model adaptive control procedures.IEEE Trans Biomed Eng 1986, 34, 567.
[13]Zadeh, L. Fuzzy Sets.Inf. Control 1965, 8, 338-53.
[14]Mamdani, E. Application of fuzzy algorithms for control of simple dynamics plant. Proc. IEEE 1974, 121, 585-8.
[15]Mamdani, E. Twenty years of fuzzy control: experiences gained and lessons learnt.Proceedings of the Second IEEE International Conference on Fuzzy Systems 1993, 339-44.
[16]Carollo, A.; Tobar, A.; Hernandez, C. A rule-based postoperative pain controller: simulation results.Int J Biomed Comput 1993, 33, 267-76.
[17]Ying, H.; McEachern, M.; Eddleman, D. W.; Sheppard, L. C. Fuzzy control of mean arterial pressure in postsurgical patients with sodium nitroprusside infusion.`.IEEE Transactions on Biomedical Engineering 1992, 39, 1060-69.
[18]Mamdani, E. H.; Østergaard, J. J.; Lembessis, E. In Advances in fuzzy sets, possibility theory, and applications; P. P. Wang, Ed.; Plenum Press: New York, 1983.
[19]Fuzzy logic in artificial intelligence. 8th Austrian artificial intelligence conference FLAI '93. Proceedings.; Klement, E. P.; Slany, W., Ed.; Springer-Verlag: Berlin, Germany, 1993; Vol. 695.
[20]Glorennec, P. Y. In Fuzzy Logic; R. Lowen and M. Roubens, Ed.; Kluwer Academic Publishers: Dordrecht, The Netherlands, 1993; pp 541-551.
[21]Nauck, D.; Klawonn, F.; Kruse, R. Combining neural networks and fuzzy controllers In: Fuzzy Logic in Artificial Intelligence. 8th Austrian Artifical intelligence Conference, FLAI '93. E. P. Klement and W. Slany. Linz, Austria: Springer-Verlag, 1993:35-46.
[22]Takagi, H.; Lee, M. Neural networks and genetic algorithm approaches to auto-design of fuzzy systems. In: Fuzzy Logic in Artificial Intelligence. 8th Austrian Artifical intelligence Conference, FLAI '93. E. P. Klement and W. Slany. Linz, Austria: Springer-Verlag, 1993:68-79.
[23]Nowé, A.; Vepa, R. A reinforcement learning algorithm based on 'safety' In: Fuzzy Logic in Artificial Intelligence. 8th Austrian Artifical intelligence Conference, FLAI '93. E. P. Klement and W. Slany. Linz, Austria: Springer-Verlag, 1993:47-58.