INTRODUCTION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Weaning a patient from mechanical ventilation means, strictly speaking, "a slow decrease in the amount of ventilator support with the patient gradually assuming a greater proportion of overall ventilation" (1), although it is often taken to mean any method of discontinuing mechanical ventilation (2). In any case, a very significant fraction of a patient's time in the intensive care unit (ICU) is typically taken up with weaning (3). Currently, weaning tends to be undertaken in a somewhat ad hoc manner, its course in any particular case being dictated by the experience and intuition of the attending physician. However, there is evidence that weaning may proceed more efficiently if directed according to some specified protocol (4, 5). Indeed, previous attempts have been made to formulate the weaning process as an algorithm (6), such as could be automated on a computer. This has the potential to alleviate pressure on medical personnel in the ICU and permit much more frequent assessment and adjustment of ventilatory parameters than the current typical rate of once or twice per day (11).

There are several ventilatory modes that have been used for weaning, and some approaches indicate that regular trials of spontaneous breathing hasten the weaning process (12). However, key to any of these approaches is the notion that a decision has to be made periodically about whether to increase or decrease the level of ventilatory support, or even whether to change the mode of support. Thus, any attempt to automate the weaning process must first deal with the problem of automatically controlling the level of ventilatory support provided to a patient. This is essentially an exercise in feedback control, and it is currently performed by a physician who regularly monitors a patient's ability to cope with a given support level and then decides whether support should be further reduced, maintained constant, or increased. Obviously, any algorithm that attempts to codify ventilatory adjustment must also utilize feedback control.

The purpose of the present study was to develop a fuzzy logic method for automatically controlling the level of pressure support ventilation (PSV) given to patients with severe chronic obstructive pulmonary disease (COPD). We decided to use fuzzy logic because it naturally allows for control algorithms to be based on human experience rather than on deterministic mathematical models.

	METHODS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Our fuzzy logic weaning algorithm uses the information contained in four vital sign parameters in order to arrive at its weaning decisions. These four parameters are heart rate (HR), arterial oxygen saturation (Sa_O2), tidal volume (VT), and breathing frequency (RR). The algorithm considers both the values of these parameters at the time a decision is to be made, as well as their rates of change. The values of the parameters are used to arrive at a characterization of the patient's current condition, and the rates of change are used to decide on the trend in this condition. Both current condition and trend are then used to decide if ventilatory support should be altered and by how much.

Fuzzification of the Parameters

The first step in the development of the fuzzy logic weaning algorithm is to fuzzify the parameters HR, Sa_O2, VT, and RR. Each parameter has a so-called range of discourse which is partitioned into a number of overlapping fuzzy sets. The complexity of the fuzzy algorithm increases dramatically with the number of fuzzy sets, so we have tried to keep the number of sets for each parameter as small as possible (we use two sets for Sa_O2, three for HR, and four each for RR and VT). Each fuzzy set has an amplitude associated with every point in its range that varies between 0 and 1, depending on how strongly a particular point in the range is considered to belong to that set. How we decided to fuzzify the four parameters specifically for patients with severe COPD is shown in Figure 1. Consider, for example, the three fuzzy sets for HR shown in Figure 1. The three sets are labeled HIGH, OK, and LOW. The OK set is chosen by first identifying that range of HR that is definitely normal and assigning it a membership value of 1.0 (in the example in Figure 1 this range runs from 70 to 100 beats/min, indicated by points b and c). Next, we identify those values of HR that define the absolute extremes of the OK range, which means that anything outside this range is definitely not OK. In the example shown in Figure 1 the lowest and highest possible OK values for HR are 50 and 150 beats/min, respectively (points a and d). Membership values at the edges of the NORMAL set are then made to decrease linearly from 1.0 at the extremes of the definitely OK range (points b and c) down to zero at the points demarcating definitely not OK (points a and d).

View larger version (36K): [in this window] [in a new window]   Figure 1.   Fuzzification of HR, SaO2, VT, and RR. The allowable ranges of each parameters are divided into overlapping fuzzy sets, each of which indicates membership over the range 0 to 1. The points a, b, c, and d indicated on the HR sets are explained in the text. The hatched areas indicate emergency regions where immediate medical intervention is required.

The HIGH fuzzy set overlaps the right-hand edge of the OK set only where membership in the OK set is less than 1.0. Specifically, at the point in Figure 1 where OK membership starts to decrease from 1.0 (point c), membership in the HIGH set starts to increase from 0 and proceeds linearly up to a value of 1.0 at the point where the OK membership reaches 0 (point d). Thus, there is a range of HR values (between points c and d in Figure 1) within which the patient might possibly be either OK or HIGH. When HR reaches values that are definitely high (at point d), they cannot also be considered possibly OK, so at this point the HIGH membership reaches 1.0 and the OK membership reaches 0. The LOW fuzzy set is constructed in an analogous way. A specific value of HR may thus be definitely OK, HIGH, or LOW if it has membership of 1.0 in any of the corresponding fuzzy sets. Alternatively, if the HR value falls into one of the overlapping regions it is considered, for example, "possibly normal" but also "possibly high" with relative likelihoods given by its respective memberships in the OK and HIGH sets. Any value of HR that falls outside the range of values encompassed by all three fuzzy sets (that is, below 40 and above 150 beats/min) is considered to represent an emergency situation requiring immediate medical intervention.

The process of dividing the range of a variable into overlapping sets and defining membership as illustrated above is called fuzzification (15). Of course, an experienced physician might well disagree with the particular fuzzifications we have chosen for patients with severe COPD (Figure 1), and might choose to locate the various fuzzy sets somewhat differently. This, however, merely illustrates the power and flexibility of the fuzzification process. It encapsulates the expertise of the person doing the fuzzifying, but without their having to explicitly formulate their thoughts. This is highly reminiscent of the subjective decision-making process that characterizes so much of human endeavor, and medical practice in particular.

Determining Patient Condition

The various fuzzy sets for HR, Sa_O2, VT, and RR define a global measure of the patient's status, which is assigned to a quantity called CONDITION having four categories: POOR, QUESTIONABLE, MODERATE, and GOOD. Every combination of set memberships for HR, Sa_O2, VT, and RR corresponds to one of these CONDITION categories. For example, obviously the best-case scenario one can hope for is that all four variables should be in their OK sets, so this combination of memberships would correspond to a CONDITION category of GOOD. Similarly, if all four variables are in their lowest sets then the CONDITION category is clearly POOR. The complete rule table for determining CONDITION used in the present study is shown in Table 1. This is not, of course, the only possible table one could construct, and an expert reader might disagree with some of the entries in it. It therefore encapsulates the collective wisdom of the investigators and is not meant to be an indication of absolute truth.

Assessment of vital signs in a patient produces a particular set of values for HR, Sa_O2, VT, and RR. Each parameter will have finite membership in either one or two of its fuzzy sets, producing between four and 16 unique set membership combinations. The lowest membership value in each combination determines the membership value in its corresponding CONDITION category (Table 1). Thus, for example, if HR, Sa_O2, VT, and RR had memberships in their respective OK sets of 1, 1, 0.8, and 0.95, then they would together constitute a membership of 0.8 in the CONDITION category GOOD. Because any particular invocation of the CONDITION rules can produce up to 16 combinations, some of the CONDITION categories may be contributed to more than once. The final membership value of a CONDITION category is chosen to be the largest of all its contributors. The overall assessment of the patient's current condition is embodied in the membership levels of the four CONDITION categories.

Determining Patient Trend

Our fuzzy logic algorithm also needs to consider whether a patient's condition is stable, improving, or deteriorating. In general, this requires the simultaneous consideration of the current values of the four parameters as well as their respective rates of change. However, treating all these eight quantities in a single fuzzy logic structure would lead to an impractically large rule table. We therefore use subsets of the parameters and their rates of change to define patient trend. Specifically, we consider RR and the rates of change of RR and Sa_O2 (labeled RR and Sa_O2, respectively) to arrive at a quantity called TREND, which is composed of four categories labeled STABLE, IMPROVING, DETERIORATING, and CRASHING. RR and Sa_O2 are fuzzified, as shown in Figure 1. The rule for determining the categories of TREND from the fuzzy set combinations for RR, RR, and Sa_O2 are given in Table 2.

Calculating the Control

At this point, we have obtained a global assessment of patient status in terms of partial memberships in four CONDITION categories, together with assessments of how status is changing in terms of the categories of TREND. Next, we have to translate these assessments into a specific action to be taken. There are three distinct quantities whose adjustments typically must be considered for the weaning patient: (1) the rate of synchronized intermittent mandatory ventilation (SIMV), (2) the level of PSV, and (3) the inspired oxygen fraction (FI_O2). The control of all three quantities simultaneously can, in principle, be undertaken by our fuzzy logic approach, but the situation is complicated significantly by the fact that they are not independent of each other. Therefore, as PSV has been shown to be particularly suited to the weaning phase of an ICU patient's ventilatory care (13, 14), in the present study we consider only the automatic adjustment of PSV level by fuzzy logic.

PSV level cannot be less than zero or greater than 25 cm H₂O. Therefore, changes in PSV level must be calculated with respect to these fixed limits. Furthermore, the actual magnitude of a change in PSV level bears a nonlinear relationship to its qualitative characterization, depending on the current PSV level. For example, if the current PSV level is 20 cm H₂O then a small increase might be 1 cm H₂O and a large increase 4 cm H₂O. However, for a current level of only 5 cm H₂O a small change might be 5 cm H₂O and a large change 15 cm H₂O. We thus define a change in PSV level in terms of the difference between the current value and the appropriate limiting value. For example, an increase of 50% would mean adding 5 cm H₂O to a current value of 15 cm H₂O, or 7 cm H₂O to a current value of 11 cm H₂O (given that the ceiling is 25 cm H₂O). Conversely, a decrease of 50% from the same two current values would mean absolute decreases of 7.5 and 5.5 cm H₂O, respectively.

Four changes in percent PSV level (labeled %PSV-CHANGE) were defined and labeled DECREASE, MAINTAIN, INCREASE, and INCREASE A LOT. Their values were 25%, 0, 25%, and 50%, respectively. Each combination of memberships in the CONDITION and TREND categories corresponds to membership in one of the %PSV-CHANGE levels, according to the rules given in Table 3. In any particular situation, there will likely be finite membership in several of the %PSV-CHANGE categories, with each category receiving a contribution from one or more combinations of CONDITION and TREND categories. When a %PSV-CHANGE category receives multiple contributions, its final membership level is determined by the largest of the contributors.

The final step in the process is to arrive at a precise value for the actual percent change in PSV level to be implemented (the so-called "crisp" control). This is done by calculating the weighted average of the memberships in each of the four %PSV-CHANGE categories. Thus, for example, a membership of 0.5 in INCREASE and a membership of 0.5 in INCREASE A LOT would result in a final crisp control of 0.5 × 25% + 0.5 × 50% = 37.5%. This number is then rounded to the nearest integer to give the amount (in cm H₂O) by which the PSV level should be altered.

The entire fuzzy logic algorithm is summarized in Figure 2.

View larger version (21K): [in this window] [in a new window]   Figure 2.   The overall structure of the fuzzy logic algorithm for controlling pressure support level. Each combination of fuzzy set memberships for VT, RR, SaO2, and HR corresponds to membership in one of four CONDITION categories via a rule table (Table 1). Similarly, fuzzy set measurements for RR, SaO2, and the rates of change of both parameters correspond to membership in four TREND categories (Table 2). Finally, each combination of categories for CONDITION and TREND determine membership four levels of %PSV-CHANGE through another rule table (Table 3). The final %PSV-CHANGE that is actually implemented is determined as the membership-weighted sum of the four %PSV-CHANGE levels.

Patient Studies

We tested our fuzzy PSV control algorithm retrospectively on 13 patients with severe chronic COPD who had been admitted for pneumonia and were undergoing a weaning trial in the intensive care unit (ICU) of the Montreal Chest Hospital. The patients (six male and seven female 70.4 ± 10.5 yr of age and weighing 69.5 ± 11.2 kg) were intubated and mechanically ventilated for respiratory failure. They had a mean ± SD stay in the ICU of the Chest Hospital of 9.3 ± 9.4 d. These patients were generally "difficult" cases and had had mean ± SD stays in other hospital ICUs of 19.2 ± 12.1 d prior to being transferred to the Chest Hospital. The patients received PSV ventilation from the Puritan Bennett 7200 ventilator (Puritan-Bennett, Carlsbad, CA), which also provided measurements of RR and VT. HR and Sa_O2 were provided by the SpaceLabs Patient Care Monitor 90303A (SpaceLabs, Medilogic, Rexdale, ON, Canada). HR, RR, VT, and Sa_O2 were obtained every hour by the attending nurses and recorded on the patients' medical charts. The patients were all eventually successfully weaned, at which point they were awake, alert, clinically stable, and able to breathe spontaneously. Access to patient records was approved by the local hospital ethics board.

We obtained the values of HR, RR, VT, and Sa_O2 from the charts for each patient over two study periods: the first 24 h the patients were admitted to the Chest Hospital ICU, and the last 24 h prior to extubation. From these values we calculated CONDITION every hour according to our fuzzy algorithm described above. The differences between successive values of RR and Sa_O2 provided the rates of change necessary to determine TREND. We thus obtained a recommended change in PSV level every hour.

	RESULTS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Histograms of the differences between the changes in PSV level suggested by our fuzzy algorithm and those actually made for the two 24-h study periods are shown in Figure 3. (It should be noted that attending physicians made changes in PSV level at most once or twice per day during rounds, whereas the algorithm made change recommendations each hour.) In both cases in Figure 3, the peak of the histogram occurs at zero, which corresponds to precise agreement between the fuzzy logic algorithm and the actual state of affairs. For the first 24-h period, there was 78% agreement to within ± 2 cm H₂O between the fuzzy logic algorithm and what actually transpired, and 91% agreement to within ± 4 cm H₂O. For the last 24-h period, there was 72% agreement to within ± 2 cm H₂O between the fuzzy logic algorithm and what actually transpired, and 84% agreement to within ± 4 cm H₂O.

View larger version (24K): [in this window] [in a new window]   Figure 3.   Histograms of the differences between the suggested changes in pressure support level provided by the fuzzy logic algorithm and what actually transpired, for all time points and all patients studied. A and B correspond to the first and last 24-h study periods, respectively.

The time course of PSV level averaged for all patients studied, together with the corresponding fuzzy algorithm recommendations, for both 24-h study periods are shown in Figure 4. In all cases the mean agreement was within 2 cm H₂O.

View larger version (49K): [in this window] [in a new window]   Figure 4.   The mean pressure support levels at each hourly time point for all patients studied (solid bars) together with the mean suggestions made by the fuzzy logic algorithm (open bars). A and B correspond to the first and last 24-h study periods, respectively.

	DISCUSSION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

The basic assumption in any weaning algorithm is that it is possible to measure a number of parameters that together contain sufficient information to allow a sensible weaning decision to be made. In the case of the present study, we assume that this information is contained in the four parameters RR, VT, HR, and Sa_O2. These are not the only possibilities, of course. East and coworkers (8), for example, argued that intrinsic positive end-expiratory pressure and FI_O2 can be used to control ventilatory adjustment, whereas Strickland and Hasson (10) utilized respiratory rate, minute ventilation, and Sa_O2 as inputs to a computer-controlled weaning system. We decided to use RR, VT, HR, and Sa_O2 because they have obvious physiological relevance to the weaning problem and because all four quantities can be measured easily, noninvasively, and often. It is impossible to say a priori, however, whether these four parameters (or, indeed, any collection of such quantities) contain enough information to successfully guide weaning. The only way to effectively test such an algorithm is by applying it to real patients and gauging weaning success.

Before embarking on such an undertaking, however, one must have at least some idea that the weaning algorithm will perform in an acceptable manner and not pursue a weaning course that is completely ridiculous. This is particularly true in the case of our algorithm because fuzzy logic has never been applied in the area of weaning before. Thus, our aim in the present study was to obtain a preliminary indication of how our fuzzy logic algorithm performs when applied to the ventilatory adjustment of real patients. The results clearly indicate that our fuzzy algorithm for pressure support adjustment makes recommendations that are generally compatible with those actually implemented in patients (Figures 3 and 4), with agreement occurring in a large fraction of cases. In those cases in which the algorithm did not agree with what actually transpired, the difference in recommended changes in PSV level were mostly small. Disagreements of this nature would be expected even between different physicians because there is still no universally accepted method for undertaking the general weaning process (1). Indeed, the conclusions of the fuzzy algorithm depend greatly on whose expertise it encapsulates (i.e., who fuzzified the parameters and set up the rule tables). Thus, we would expect some disagreement even between different realizations of the fuzzy algorithm produced by different experts. Also, the vital signs reported in the patients' charts were single values recorded by the nursing staff. Whereas one might expect Sa_O2 to remain reasonably stable over a time span of minutes, RR, VT, and HR can vary considerably within a few seconds, so we cannot be sure that the recorded values were a fair representation of the averages of these parameters at the time of recording. More representative parameter values would be obtained if the parameters could be recorded continuously on a computer for a minute or so prior to each assessment point, and averages taken. However, this facility is not available in most ICUs, including that of the Montreal Chest Hospital where our studies were performed.

The attending physicians treating the patients in our study rounded typically once or twice per day. Consequently, most of the hourly assessments that our fuzzy logic algorithm made amounted to questioning the status quo, so the agreement shown in Figure 3 needs to be interpreted with some caution. Nevertheless, this does not invalidate the test of the algorithm because we still need to establish that those fluctuations that did occur in the patients' vital signs from hour to hour would not cause the algorithm to produce unreasonable recommendations. As a more stringent test of our algorithm, we were able to identify in some cases from the patient charts that a change in PSV level had been made at a certain time by the attending physician. These instances therefore allowed us to make a direct comparison between the decision of a physician and the recommendation of the fuzzy algorithm. These comparisons, which demonstrate that in some cases the agreement between physician and fuzzy algorithm was good, are shown in Figure 5 (good agreement is shown by line segments that do not end far from the line of identity). In other cases, however, the agreement was rather poor (poor agreement is shown in Figure 5 by line segments ending far from the line of identity). The reasons for these cases of poor agreement were varied, and not always clear. For example, in one case we found that the physician and fuzzy algorithm agreed well until the physician implemented a large decrease in PSV level, only to return it to the original level some hours later. The fuzzy algorithm, in contrast, maintained a constant recommendation throughout, so we presumed that the physician had experimented with a maneuver that ultimately turned out to be unsuccessful. Thus, a number of the specific instances involving physician intervention seemed to indicate actions that did not reflect the application of a uniform set of PSV adjustment criteria. Interestingly, the general agreement between physician and fuzzy algorithm appeared better in the first 24-h period (Figure 5A) than in the last 24-h period (Figure 5B). In particular, in the last 24-h period, the physician reduced PSV levels in a more aggressive manner than the algorithm would have, which may have reflected some intuition on the part of the physician that weaning was imminent. However, more aggressive support reduction is appropriate from the physicians because they only check and modify PSV levels infrequently. Our fuzzy logic algorithm is potentially able to make frequent small modifications to PSV level, say every hour throughout the day, that could amount to a similar overall change as a single physician-directed modification implemented once a day.

View larger version (18K): [in this window] [in a new window]   Figure 5.   Comparison between changes actually made in PSV level by an attending physician and changes recommended by the fuzzy logic algorithm. Each line segment begins (closed circles) on the diagonal line of identity to show the initial PSV level, and ends (open circles) at the position indicating corresponding physician-implemented and algorithm-recommended PSV values.

Fuzzy logic was introduced by Zadeh (16) in the 1960s and is now well established as an engineering discipline (15). It enjoys widespread use as a means for developing algorithms for controlling a wide variety of devices (7, 15) and to date has been exploited most fully in Japan (19). The principal advantage of fuzzy logic over other control paradigms is the way it so naturally represents subjective human notions. In other words, rather than always requiring definite decisions to be made in the manner demanded of algorithms based on conventional mathematical equations and rule-based systems, fuzzy logic allows for uncertainty to be built into a control algorithm in a specific fashion. Fuzzy logic has been used in applications that are also amenable to conventional control algorithms based on mathematical models of the system being controlled such as the high-frequency mechanical ventilator of Noshiro and coworkers (20). However, fuzzy logic has a particular advantage in areas where precise mathematical descriptions of the control process are impossible. Such an area is medical decision-making involving a variety of factors of varying relative importances that are weighed up against a background of experience. In the case of weaning from mechanical ventilation, the only algorithms that have been published previously are of the conventional or "recipe" type such as the complicated logic of East and coworkers (8) and the knowledge-based system of Dojat and coworkers (7). We feel, however, that weaning is perfectly suited to the fuzzy approach because of its heuristic nature.

We must stress that the fuzzy algorithm we have developed in this study is by no means the only one possible. We have limited ourselves to considering only the four parameters RR, VT, HR, and Sa_O2, and we have also utilized a logic structure that is a small subset of that possible even with these four quantities. This was done to reduce the number of entries required in the rule tables. By contrast, we could have chosen to estimate an overall patient trend using all four parameters and their rates of change, but this would have involved an unmanageably large rule table with 4 × 4 × 3 × 2 × 3 × 3 × 3 × 3 = 7,776 entries. We thus calculated two trends, each using a subset of the possible parameters and their rates of change, for a total of 45 table entries. However, we do not claim that this is the only valid way of reducing the complexity of the algorithm.

Although we have developed a fuzzy algorithm that only adjusts pressure support level, the fuzzy approach can be extended in a straightforward manner to deal with other aspects of weaning such as adjustment of SIMV rate and FI_O2. These might require additional inputs such as the current FI_O2, but the principle is the same. Indeed, we envisage the ultimate development of a completely automatic weaning algorithm. An intriguing aspect of such an algorithm is that it would encapsulate the expertise of the medical personnel who defined the fuzzy sets and rule table entries. Thus, by having different ICU physicians establish their own fuzzy algorithms, one could subject each algorithm to a range of patient scenarios and so have a means of quantitatively assessing the degree of consensus between the physicians. At present, the only means of comparing different weaning protocols is through expensive clinical trials (12).

In conclusion, we have developed an algorithm for adjustment of pressure support mechanical ventilation using measurements of RR, VT, HR, and Sa_O2 and based on fuzzy logic. The algorithm calculates assessments of patient status and how that status is changing, and decides how to adjust pressure support level accordingly. We retrospectively compared the decisions of our algorithm in 13 ICU patients against what was actually implemented, and found generally good agreement. Although most of these comparisons were made against situations in which no medical assessment was made, these results show that our algorithm did not behave erratically and provided generally stable recommendations. We conclude that fuzzy logic provides an effective, convenient, and natural means of codifying the medical decisions made during adjustment of PSV level, and that it has promise as a means of ultimately automating the entire weaning process.