Share this article on:

Applying Computer Models to Realize Closed-Loop Neonatal Oxygen Therapy

Morozoff, Edmund MASc; Smyth, John A. LRCPSI, FRCPC; Saif, Mehrdad PhD

doi: 10.1213/ANE.0000000000001367
Technology, Computing, and Simulation: Original Clinical Research Report

BACKGROUND: Within the context of automating neonatal oxygen therapy, this article describes the transformation of an idea verified by a computer model into a device actuated by a computer model. Computer modeling of an entire neonatal oxygen therapy system can facilitate the development of closed-loop control algorithms by providing a verification platform and speeding up algorithm development.

METHODS: In this article, we present a method of mathematically modeling the system’s components: the oxygen transport within the patient, the oxygen blender, the controller, and the pulse oximeter. Furthermore, within the constraints of engineering a product, an idealized model of the neonatal oxygen transport component may be integrated effectively into the control algorithm of a device, referred to as the adaptive model. Manual and closed-loop oxygen therapy performance were defined in this article by 3 criteria in the following order of importance: percent duration of SpO2 spent in normoxemia (target SpO2 ± 2.5%), hypoxemia (less than normoxemia), and hyperoxemia (more than normoxemia); number of 60-second periods <85% SpO2 and >95% SpO2; and number of manual adjustments.

RESULTS: Results from a clinical evaluation that compared the performance of 3 closed-loop control algorithms (state machine, proportional-integral-differential, and adaptive model) with manual oxygen therapy on 7 low-birth-weight ventilated preterm babies, are presented. Compared with manual therapy, all closed-loop control algorithms significantly increased the patients’ duration in normoxemia and reduced hyperoxemia (P < 0.05). The number of manual adjustments was also significantly reduced by all of the closed-loop control algorithms (P < 0.05).

CONCLUSIONS: Although the performance of the 3 control algorithms was equivalent, it is suggested that the adaptive model, with its ease of use, may have the best utility.

Published ahead of print August 2, 2016.

From the *British Columbia’s Women’s Hospital and Health Center, Vancouver, British Columbia, Canada; School of Engineering, Simon Fraser University, Burnaby, British Columbia, Canada; and Faculty of Engineering, University of Windsor, Windsor, Ontario, Canada.

Published ahead of print August 2, 2016.

Edmund Morozoff, MASc, is currently affiliated with the New Product Development, Philips Respironics, Kennesaw, Georgia.

Accepted for publication March 24, 2016.

Funding: Clinical evaluation funding was provided by the Vancouver Foundation.

The authors declare no conflicts of interest.

This report was previously presented, in part, at the Innovations and Applications of Monitoring Perfusion, Oxygenation and Ventilation (IAMPOV) Symposium 2015. This is the second submission. It has been revised and submitted as a research report.

Reprints will not be available from the authors.

Address correspondence to Edmund Morozoff, MASc, New Product Development, Philips Respironics, 175 Chastain Meadows Court, Kennesaw, GA 30144, Address e-mail to paul.morozoff@philips.com.

Oxygen therapy is the process of administrating elevated levels of fractional inspired oxygen (FiO2) as a means of maintaining adequate tissue oxygenation. Preterm babies often require continuous oxygen therapy as a result of their diseased or underdeveloped lungs. The goal of this therapy is to provide enough oxygen to maintain normal physiologic function and minimize the risk of oxygen toxicity.

Prolonged durations of high tissue oxygen levels (hyperoxemia) have been associated with damage to the retina (retinopathy of prematurity)1 and the development of chronic lung disease (bronchopulmonary dysplasia)2; however, low levels of oxygen (hypoxemia) may be damaging and potentially mortal.3 The changes in the neonatal physiologic condition involving lung ventilation and perfusion, composition of hemoglobin (fetal versus adult), and oxygen affinity of the blood can make the achievement of safe levels of tissue oxygenation challenging.4,5

Adequate tissue oxygenation is the goal of oxygen therapy, and arterial oxygen saturation (SaO2) is its key indicator. Figure 1 illustrates the key components of the manual oxygen therapy process and their associated relationships. Nursery staff members use a pulse oximeter value (SpO2) as a proxy measure of SaO2 and manually adjust FiO2 levels via an air/oxygen blender to maintain SpO2 levels within a specific range. As shown in Figure 1, SpO2 is not the only input that they use as part of their decision making process, and FiO2 may not be the only output associated with maintaining safe levels of oxygen.

Figure 1

Figure 1

For the aforementioned reasons, neonatal oxygen therapy is difficult, and a need for accurate full-time oxygen therapy has been identified.6 Two models of the oxygen transport system are presented in this article. One is part of the oxygen therapy system model and is used to develop, verify, and compare control algorithms. The second model, an idealized version of the first one, is part of the control algorithm of an automatic oxygen control device. Results from a clinical study in which we compared the performance of the automatic closed-loop oxygen therapy device to manual oxygen therapy are presented and discussed. This article demonstrates that computer modeling is not only a vehicle for facilitating closed-loop control development but also that an idealized version of the model can become part of the realized controller’s algorithm.

Back to Top | Article Outline

METHODS

This section describes the development of the computer model and its integration of the neonatal component into an electromechanical device. It also describes a clinical study that compared the performance of the model-based algorithm and 2 other control algorithms with manual oxygen therapy. The study was performed at British Columbia’s Children’s Hospital and was approved by their associated IRB before it began. Written informed consent was obtained from the parents of all babies before entering them into the study. After study approval, a presentation was given to nursery staff members that described the study and demonstrated how the controller operated.

Back to Top | Article Outline

Automating Neonatal Oxygen Therapy

In Severinghaus and Astrup’s7 historical review of pulse oximeters, they described the early 1940s work of Millikan and Pappenheimer, who were tasked with the challenge of preventing fighter pilots from passing out during high-altitude combat. Researchers Beddis et al8 and Collins et al9 believed that there was a need to maintain adequate PaO2 levels in preterm infants. The device they developed in the 1970s, known as Twiggy, was the first closed-loop neonatal oxygen therapy device.

In the 1980s and 1990s, with the development of the pulse oximeter, researchers turned their attention toward using SpO2 as the input for automatic oxygen therapy. At this time, work was also being done, with computer models, to develop and verify various control strategies.10–12 Taube13 was the first researcher to develop an automated oxygen therapy device that had SpO2 as the input. This device was designed for incubated, nonventilated babies. In the early 1990s, Morozoff et al14,15 developed and studied an SpO2 controller for use on ventilated very low-birth-weight babies. Researchers continued to refine and improve the technology,16–21 and by 2010, the first commercial closed-loop oxygen therapy device was released for sale. Claure and Bancalari22 provide a comprehensive summary of automated oxygen therapy clinical results in their 2013 article.

Back to Top | Article Outline

Computer Models

Modeling the physiologic system provides several advantages toward automating neonatal oxygen therapy. It decreases development times and reduces patient risk. Clinical studies are challenging and can hinder timely device validation, whereas a model can provide a means of initially verifying equipment before committing to a clinical evaluation or study. Another benefit is that the model itself can be integrated into the control algorithm, becoming part of the algorithm: observing, learning, and providing additional inputs toward maintaining adequate SpO2 levels.

Gray23 proposed the first quantitative model of human ventilation. His steady-state model described ventilation rate, in a feedback control loop, controlled by PacO2, PaO2, and acidemia. Dynamic models were evolved by Grodins et al24 and Defares et al.25 Computer-based respiratory models were first explored by Horgan and Lange26 as well as Milhorn and Guyton.27 Later, Yamamoto and Hori28 introduced bidirectional flow using a periodic respiratory drive. Many of these researchers modeled the respiratory system at various levels of detail based on the resistance-compliance (RC) network analogy described by Grodins29 and Milhorn.30 Rideout31 produced a model that has been adopted and modified for a neonate model by the author.32

Four components make up the model view of the closed-loop system illustrated in Figure 2. Each is executed within a single software program that manages the movement of inputs/outputs, actual FiO2, SaO2, and SpO2. The controller has knowledge of the system response and can maintain a history of its performance. It applies this knowledge to the SpO2 value it receives from the pulse oximeter model and determines whether a change in FiO2 is warranted. A model of the air/oxygen blender adjusts the delivered FiO2 to the neonate model. The neonate model responds to FiO2 changes and delivers new SaO2 values to the pulse oximeter. A script within the program can initiate shunts within the neonate model creating sudden saturation changes, as one may see from a labile neonate.

Figure 2

Figure 2

From the authors’ personal experience in the neonatal clinic and discussions with clinical staff, shunting as a result of blood bypassing the lungs or poor perfusion in areas of the lungs creates the most sudden and dramatic changes in neonatal SaO2 levels. Although there are other factors that can affect SaO2 levels, such as adult versus fetal hemoglobin, H+ (acidity), body temperature, etc, these factors tend to provide relatively slow changes in tissue oxygenation compared with shunting conditions. Slow changes are more manageable by control systems, whereas sudden disturbances can create instability or poor response in some controllers.

The reader should note that the computer model in Figure 2 contains only the inputs and outputs of the outer loop of Figure 1. This simplification is driven by the desire to simplify the controller. The addition of each sensor to a product increases additional development time, product cost, potential failure modes, and complexity. Furthermore, a more complex model generally requires more system memory, computing power, and user expertise. When designing a product, engineers must weigh the benefits of potentially increasing the performance of a control system by adding more inputs and outputs against the risk of adding more failure modes and further complexity. The goal is not only to develop a verified and validated product that is safe and effective but also to produce a product within a reasonable time frame, cost, and ease of use. With these points in mind, the oxygen therapy process and its associated model have been idealized into the model shown in Figure 2.

Back to Top | Article Outline

Modeling the Neonate

The oxygen transport system consists of 2 primary models, the respiratory and circulation systems linked by the Severinghuas simplified oxygen dissociation equation33 at the blood-gas interface within the alveoli as shown in Figure 3.

Figure 3

Figure 3

Gas flow through a compartment is driven by a pressure differential and is also a function of that compartment’s RC as illustrated in Figure 4. The pressures, flows, and volumes can be can be represented by a series of mathematical equations. This fundamental building block is used to create the multiple compartments of the respiratory model. In parallel with each compartment, the pressures, flows, and volume values for oxygen can also be derived as shown in Figure 5. The output of the model during inspiration, PaO2, represents the partial pressure of oxygen in the alveoli. This value is used as an estimate of PaO2 in the blood in the pulmonary capillaries, an input into the oxygen dissociation curve. PaO2, as shown in Figure 5, also represents the partial pressure of oxygen in the alveoli during expiration and is derived from SvO2 within the pulmonary artery.

Figure 4

Figure 4

Figure 5

Figure 5

Within the blood, oxygen bonds to hemoglobin as a function of PaO2 per the relationship defined by the oxygen dissociation curve. Because blood, like most fluids, is noncompressible, the compartments of the cardiovascular system can be treated as mixing chambers. We can model the transport of oxygen within the cardiovascular system as a solute with concentrations, SaO2 and SvO2, carried by the blood.

The output concentration of a material from a mixing chamber is a function of the solute carrier flow, chamber volume, and input concentration.34 Three types of chambers are modeled: a simple mixing chamber used for most compartments; connecting chambers used for capillaries; and consumption chamber used for tissue.

A series of these compartments can be linked together to create a larger model of the cardiovascular system as illustrated in Figure 6. This model is based on the work of Rideout’s 10-compartment model31 and the assumptions that blood flow is nonpulsatile and unidirectional, all mixing chambers are perfect, and red blood cells travel at the same rate as the blood.

Figure 6

Figure 6

In summary, it should be noted that the oxygen transport system is complex, with many inputs and outputs, and a model could eventually be developed that is just as complex. However, from an engineering standpoint, simplifying the model provided us not with perfection but with sufficient guidance.

A complete, idealized adult model was first developed on a personal computer as a means of verifying the performance against other published results.31 This model was then modified to represent a neonate’s physiology by adjusting physiologic variables such as the respiratory compartment RC, tidal volume, cardiovascular compartment volumes, blood flow, transport time, and cardiac output. The variables were based on values available in the literature and allometric equations scaled for a neonate weighing 1500 g. The neonate model software was then run on a PC workstation and its output compared with published data. The model was then used to experiment with fuzzy logic control and motion artifact rejection without the need for initial animal or human studies.35 Computer modeling of neonatal oxygen transport is realizable and can be a useful tool for creating control algorithms.

Back to Top | Article Outline

Integrating the Model Into the Algorithm

It is possible to integrate the neonate model into the control algorithm. However, this migration requires that it operate in a real-time environment. As a means of meeting the product’s computing performance criteria: reduced complexity, system resources, and processing time, the model is distilled into a nonlinear transfer function; a simplified version is shown in Figure 7. This transfer function determines the relative FiO2 change required to reach a specific SpO2 target. For example, if the target is 92% and the current SpO2 value is 94%, the control algorithm, using the thick line, would reduce the current FiO2 value by 5%.

Figure 7

Figure 7

Self-tuning can be added to the model-based algorithm by creating a supervisor that tracks inputs such as the models performance (eg, percent time close to target), number of manual adjustments, number or rapid desaturations, etc. Every second the supervisor binned the percentage of time that the neonate stayed at 1 of 3 ranges, normoxemia, hyperoxemia, and hypoxemia. At 2-minute intervals, the supervisor would adjust the model to improve performance using the following algorithm:

  1. Get 3 values: normoxemia, hyperoxemia, and hypoxemia
  2. IF mostly normoxemia THEN
    • IF hyperoxemia > hypoxemia THEN slight steepening of curve
    • ELSE slight flattening of curve
  3. ELSE IF mostly hyperoxemia THEN large steepening of curve
  4. ELSE IF mostly hypoxemic THEN large flattening of curve

For example, if the neonate has been consistently hyperoxemic, the controller would steepen the latter 2 segments of the transfer function by moving points 3 and 4 to 3’ and 4’ as shown in Figure 7. This would cause the controller to increase its FiO2 change when above target; instead of a 5% change at 94% SpO2 the controller would now affect an 8% change.

Because this control method, referred to as the adaptive model (AM) algorithm, adapts itself toward maintaining maximal normoxemia, it requires no tuning or setup by hospital staff. Like the state machine (SM) algorithm, it also accounts for the nonlinearity of the system. Disturbances to the system such as shunting and bradycardia, however, could not be recognized by the instrument without further inputs such as heart rate, SvO2, etc.

Back to Top | Article Outline

Clinical Evaluation

A second-generation automatic SpO2 controller, developed at British Columbia’s Children’s Hospital, was used in this study. It contained 3 user-selectable control algorithms: SM, proportional-integral-differential (PID), and AM. The device used the SpO2 signal from a commercial pulse oximeter as an input signal and controlled the patient’s FiO2 level by actuating a motorized mechanical air/oxygen gas blender. A manual override feature was provided in the gas blender. The SM algorithm was ported from a first-generation controller, and implementation details have been published previously.14,15 The results of the first-generation controller study were used to tune the algorithm in this study as a means of improving its performance. Implementation details on the PID algorithm can be found in previous publications.36,37 SpO2 and FiO2 (per the blender setting) were sampled once per second and stored to memory for further analysis.

Performance was defined as 3 criteria in the following order of importance: percent duration of SpO2 spent in normoxemia, hypoxemia, and hyperoxemia; number of 60-second periods <85% SpO2 and >95% SpO2; and number of manual adjustments. Normoxemia for the closed-loop control algorithms was defined as ±2.5% of their target SpO2. Manual mode normoxemia was defined within the range of 90% to 95% SpO2, the range that clinical staff attempted to maintain when performing oxygen therapy per standard hospital guidelines. Target SpO2 error is the difference between the controller’s SpO2 target and the actual SpO2; it was used as an additional performance criterion for comparing the closed-loop control algorithms with each other. It was not reasonable to use this criterion for manual mode comparison because no attempt was made by clinical staff to maintain SpO2 at a specific target SpO2. Two-tailed paired t tests were used to analyze the performance criteria with P <0.05 indicating significance.

An evaluation was defined as the application of manual and automatic control on a preterm baby on a specific day. Each evaluation (day) would have ≥1 hour of manual oxygen therapy and ≥1 hour of closed-loop therapy (SM, AM, or PID) data recorded. An attempt would also be made to apply the other algorithms, each for an hour.

Back to Top | Article Outline

RESULTS

Applying the Model to Patients

Sixteen clinical evaluations were performed on 7 low-birth-weight ventilated preterm babies. We purposely chose subjects that were relatively unstable and would challenge the controller. Demographic characteristics of the enrolled babies are shown in Table 1. An investigator was present during automatic control to provide technical support, monitor the controller’s performance, and record clinical events (suctioning, patient motion, etc). Investigators did not apply therapy to the patient.

Table 1

Table 1

Table 2

Table 2

Table 3

Table 3

Table 4

Table 4

Table 5

Table 5

Table 6

Table 6

Each clinical evaluation occurred during the 8:00 AM to 10:00 PM period of a single day. Some evaluations were done on the same baby but took place on the following days as shown in Table 2. Manual FiO2 adjustment was allowed and recorded during the automatic algorithms. Also, during the trial’s closed-loop control periods, the automatic controller was not turned off during clinical interventions such as suctioning, diaper change, physiotherapy, etc. The 3 automatic algorithms SM, PID, and AM along with manual oxygen therapy were applied to the neonates as described in Table 2. Note that during some clinical evaluations, several variations of an algorithm were run for at least a 1-hour period. The PID algorithm generally required some tuning to ascertain an optimal set of gain parameters. In an effort to avoid the effects of long-term changes in physiology, the algorithms were run consecutively. Tables 3 through 5 list the performance criteria data, comparing manual therapy to each closed-loop therapy. Additionally, Tables 6 and 7 compare the AM algorithm to the SM and PID algorithms, respectively.

Back to Top | Article Outline

DISCUSSION

When compared with manual therapy, all 3 control algorithms significantly increased the patient’s duration in normoxemia by reducing time in hyperoxemia. They did not reduce time in hypoxemia when compared with manual mode. In our 1993 study, in which we compared the SM algorithm to manual therapy, the algorithm used in that study produced a slight but significant increase in the hypoxemic duration.15 The improvements made to the SM algorithm and applied in this study successfully reduced the time at hypoxemia to a duration equivalent to that of manual mode. The AM algorithm performed well and had the greatest mean duration at normoxemia, 60%, when compared with manual therapy. As shown in Table 7, it also performed better than the PID algorithm in this criterion. With respect to duration at normoxemia, however, the SM algorithm performed better than the AM.

Table 7

Table 7

The PID algorithm demonstrated a significant improvement in reducing the number of 60-second events >95% and <85%. The SM algorithm also significantly reduced the number of 60-second events >95% SpO2. Compared with manual therapy, the AM algorithm did show improvement in the number of these events; however, it was not significant. This result was not expected because it was postulated that the nonlinear nature of the model would adjust the FiO2 by smaller increments when below target and larger increments when above target, reducing potential oscillation in labile babies. The AM algorithm may have had more of these events >95% because its median target SpO2 was 94% vs the 93% median target for the PID and SM algorithms. In other words, the AM algorithm had a target that was closer to the 95% threshold of this criterion.

All 3 control algorithms significantly reduced the number of manual adjustments. The AM algorithm demonstrated the most significant improvement when compared with manual therapy. Comparing the AM algorithm to the other closed-loop algorithm (Tables 6 and 7), there was no significant difference in performance.

Regarding usability, we found that the PID algorithm worked well when its gain factors were properly tuned. This, however, required experience with the controller, understanding its response. The SM algorithm performed well and required minor setup. We found that the AM algorithm was the most convenient algorithm because it required no setup or previous experience with the controller, demonstrating that a model of the neonate can be used to create a transfer function that, in turn, can become a viable part of a closed-loop control algorithm.

Back to Top | Article Outline

CONCLUSIONS

Driven by the requirements to engineer a product that balanced safety, efficacy, and performance, we sought to automate the difficult task of neonatal oxygen therapy by reducing the manual therapy’s set of inputs and outputs. To verify control algorithms, our computer model of the neonatal oxygen transport system further simplified the neonate’s complex physiology. We idealized the neonatal computer model and integrated it into a closed-loop oxygen therapy device. Its performances against the manual process and 2 other control algorithms were assessed in the clinical setting on ventilated low-birth-weight infants. The results showed the model-based controller, like the other 2 closed-loop control algorithms, had significantly better performance than manual therapy. When compared with the other algorithms, the AM algorithm’s performance was equivalent but not significantly better. Because the AM’s performance is similar to the other algorithms and it required no initial setup or in-process tuning, it potentially may provide a slightly better O2 therapy device.

In Joseph Conrad’s novel, The Secret Agent, his character Michaelis states, “… All idealism makes life poorer. To beautify it is to take away its character of complexity – it is to destroy it.” Conrad’s words are certainly appropriate with regard to the pursuit of knowledge, truth, and understanding. However, within the challenge of engineering a product, we found utility in simplicity.

Back to Top | Article Outline

DISCLOSURES

Name: Edmund Morozoff, MASc.

Contribution: This author helped design the study, conduct the study, analyze the data, write the manuscript, and develop the model and device.

Name: John A. Smyth, LRCPSI, FRCPC.

Contribution: This author helped design the study, conduct the study, analyze the data, and write the manuscript.

Name: Mehrdad Saif, PhD.

Contribution: This author helped design the study, analyze the data, and write the manuscript.

This manuscript was handled by: Maxime Cannesson, MD, PhD.

Back to Top | Article Outline

REFERENCES

1. Payne JW. Avery ME, Taeusch HW. Retinopathy of prematurity. Schaffer’s Diseases of the Newborn. 1984:Toronto, Ontario, Canada: WB Saunders Co. 909–14.
2. Northway WH Jr, Rosan RC, Porter DY. Pulmonary disease following respiratory therapy of hyaline-membrane disease. Bronchopulmonary dysplasia. N Engl J Med 1967;276:357–68.
3. The BOOST-II Australia and United Kingdom Collaborative Groups. Outcomes of two trials of oxygen-saturation targets in preterm infants. N Engl J Med 2016:374:749–60.
4. Tin W. Oxygen therapy: 50 years of uncertainty. Pediatrics 2002;110:615–6.
5. Askie LM, Henderson-Smart DJ, Irwig L, Simpson JM. Oxygen-saturation targets and outcomes in extremely preterm infants. N Engl J Med 2003;349:959–67.
6. Flynn JT, Bancalari E, Snyder ES, Goldberg RN, Feuer W, Cassady J, Schiffman J, Feldman HI, Bachynski B, Buckley E. A cohort study of transcutaneous oxygen tension and the incidence and severity of retinopathy of prematurity. N Engl J Med 1992;326:1050–4.
7. Severinghaus JW, Astrup PB. History of blood gas analysis. VI. Oximetry. J Clin Monit 1986;2:270–88.
8. Beddis IR, Collins P, Levy NM, Godfrey S, Silverman M. New technique for servo-control of arterial oxygen tension in preterm infants. Arch Dis Child 1979;54:278–80.
9. Collins P, Levy NM, Beddis IR, Godfrey S, Silverman M. Apparatus for the servocontrol of arterial oxygen tension in preterm infants. Med Biol Eng Comput 1979;17:449–52.
10. Sano A,, Kikuchi M. Adaptive control of arterial oxygen pressure of newborn infants under incubator oxygen treatment. Proc IEEE EMBS 1985;85:205–11.
11. Yu C, He WG, So JM, Roy R, Kaufman H, Newell JC. Improvement in arterial oxygen control using multiple-model adaptive control procedures. IEEE Trans Biomed Eng 1987;34:567–74.
12. Tehrani FT. A microcomputer oxygen control system for ventilatory therapy. Ann Biomed Eng 1992;20:547–58.
13. Taube J. Automatic control of neonatal fractional inspired oxygen. 1989Thesis, Drexel University.
14. Morozoff PE, Evans RW. Closed-loop control of SaO2 in the neonate. Biomed Instrum Technol 1992;26:117–23.
15. Morozoff EP, Evans RW, Smyth JA. Automatic control of blood oxygen saturation in premature infants. Proc IEEE Conf Control Appl 1993;1:415–20.
16. Bhutani VK, Taube JC, Antunes MJ, Delivoria-Papadopoulos M. Adaptive control of inspired oxygen delivery to the neonate. Pediatr Pulmonol 1992;14:110–7.
17. Tehrani FT, Bazar AR. A feedback controller for supplemental oxygen treatment of newborn infants: a simulation study. Med Eng Phys 1994;16:329–33.
18. Raemer DB, Ji XB, Topulos GP. FIx controller: an instrument to automatically adjust inspired oxygen fraction using feedback control from a pulse oximeter. J Clin Monit 1997;13:91–101.
19. Sun Y, Kohane IS, Stark AR. Computer-assisted adjustment of inspired oxygen concentration improves control of oxygen saturation in newborn infants requiring mechanical ventilation. J Pediatr 1997;131:754–6.
20. Claure N, Gerhardt T, Everett R, Musante G, Herrera C, Bancalari E. Closed-loop controlled inspired oxygen concentration for mechanically ventilated very low birth weight infants with frequent episodes of hypoxemia. Pediatrics 2001;107:1120–4.
21. Urschitz MS, Horn W, Seyfang A, Hallenberger A, Herberts T, Miksch S, Popow C, Müller-Hansen I, Poets CF. Automatic control of the inspired oxygen fraction in preterm infants: a randomized crossover trial. Am J Respir Crit Care Med 2004;170:1095–100.
22. Claure N, Bancalari E. Automated closed loop control of inspired oxygen concentration. Respir Care 2013;58:151–61.
23. Gray JS. The multiple factor theory of the control of respiratory ventilation. Science 1946;103:739–44.
24. Grodins FS, Gray JS, Schroeder KR, Norins AL, Jones RW. Respiratory responses to CO2 inhalation; a theoretical study of a nonlinear biological regulator. J Appl Physiol 1954;7:283–308.
25. Defares JG, Derksen HE, Duyff JW. Cerebral blood flow in the regulation of respiration. Studies in the regulation of respiration. I. Acta Physiol Pharmacol Neerl 1960;9:327–60.
26. Horgan JD, Lange RL. Digital computer simulation of respiratory response to cerebrospinal fluid PCO2 in the cat. Biophys J 1965;5:935–45.
27. Milhorn HT Jr, Guyton AC. An analog computer analysis of Cheyne-Stokes breathing. J Appl Physiol 1965:20:328–338.
28. Yamamoto W, Hori T. Phasic air movement model of respiratory regulation of carbon dioxide balance. Comput Biomed Res 1970;3:699–717.
29. Grodins FS.Control Theory and Biological Systems. 1963New York, NY: Columbia University Press.
30. Milhorn HT Jr. The Application of Control Theory to Physiological Systems. 1966Philadelphia, PA: W.B. Saunders Co.
31. Rideout VC.Mathematical and Computer Modelling of Physiological Systems. 1991New Jersey, NJ: Prentice Hall, Inc.
32. Morozoff EP,, Saif M. Oxygen therapy control of neonates—part I: a model of neonatal oxygen transport. Control Intell Syst 2008:36:227–37.
33. Severinghaus JW. Simple, accurate equations for human blood O2 dissociation computations. J Appl Physiol 1979;46:599–602.
34. Sheppard CW. The theory of the study of transfer within a multi compartment system using isotropic tracers. J Appl Physics 1948:19:70.
35. Morozoff EP, Saif M. Oxygen therapy control of neonates II: evaluating manual, PID, and fuzzy logic controller designs. Control Intell Syst 2008;36:238–49.
36. Morozoff EP, Smyth JA, Evans RW. Automatic SaO2 control using adaptive modelling. Proc Conf Canadian Med Biol Eng Soc Vancouver 1994;1:144–5.
37. Morozoff EP, Smyth JA. Evaluation of three automatic oxygen therapy control algorithms on ventilated low birth weight neonates. Conf Proc IEEE Eng Med Biol Soc 2009;2009:3079–82.
© 2017 International Anesthesia Research Society