Human physiology provides many examples of natural feedback control systems. The baroreceptors in the aortic arch continuously measure physiologic variables that reflect blood pressure and intravascular volume. Neural circuits trigger corrective responses such as increased vascular tone and heart rate when a drop in the pressure is sensed, which might occur, for instance, with a change in posture. Glucostasis in a nondiabetic patient has the serum glucose level as the state variable; insulin output and the degree of gluconeogenesis are some of the control variables involved. Sensors in the ß-cell detect rising glucose levels after a meal and trigger an outflow of insulin to reduce the levels.

Much disease management can be described as feedback control. For instance, we monitor blood pressure. If it is too high, we start with low doses of blood pressure medication and increase the dose as needed to titrate the blood pressure to normal range. We measure activated partial thromboplastin time (APTT) levels when anticoagulating patients with heparin and increase or decrease the rate of heparin infusion according to the APTT level. Diabetes management is the archetypal feedback control system. It has easily measurable state variables (for example, the blood glucose level), control variables (for example, the dose of insulin), and rules (algorithms) that tell how to adjust the control variables according to the value of the state variables. Orders for sliding-scale insulin provide a simple example of a control system: Give 2 units of insulin if the serum glucose level is between 11.1 and 13.9 mmol/L (200 and 250 mg/dL), give 4 units if the level is between 13.9 and 16.7 mmol/L (250 and 300 mg/dL), and so forth.

Computerized medical information systems can improve feedback control at many levels of the health care process. They can help to gather the state variables needed on the "sensory" side of the control loop. In intensive care, computer systems monitor blood pressure, heart rate, and ventricular rhythm. In primary care, computers remind physicians when it is time for a woman's mammogram or a diabetic patient's fundoscopic examination [2]. Computers can also suggest the specific adjustments to the control variables needed to correct an out-of-control state. For example, computers can suggest specific adjustments to the dose of heparin to reach a stated target goal (for example, an APTT between 2 and 2.5 times control). In medicine, computer-based feedback systems more often include a human (physician or nurse) in the feedback loop, but the principle is the same.

Much of diabetic care depends on relatively few laboratory variables (for example, serum glucose and glycosylated hemoglobin levels) combined with patient responses to a few questions (for example, their level of exercise and the time and amount of meals). Much of this information is already available to computers from laboratory systems and home glucometers. For these reasons, diabetes is especially well suited to computer-assisted feedback control.

Closed-Loop Systems (Where the Computer Does It All)

Closed-loop control systems have great potential for improving health outcomes and reducing costs. Such systems measure the state variables directly with sensors, "decide" what dose of the control variable to "prescribe," and deliver it automatically. The typical closed-loop control system measures and adjusts the control variables continuously—like the computerized fuel injectors in a modern automobile. The protocols in closed-loop medical systems typically consist of specific equations that relate state variables to the needed changes in control variables.

Closed-loop control systems are not new to medicine. Sheppard and Kouchoukos [3] pioneered their use in the 1970s. Their system automatically infused fluids postoperatively according to the patient's pulse, hemoglobin, and blood pressure levels, and it shortened postoperative intensive care time. Other closed-loop control systems infuse nitroprusside [4, 5] and oxytocin [6] to provide faster results, with less medicine and fewer overshoots than manually controlled infusion of the same medication.

Implantable defibrillators are a successful example of closed-loop control. Like cardiac pacemakers, they are implanted in the chest, monitor the patient's cardiac rhythm, and can provide electrical shocks to control specific aberrant rhythms. However, they can deliver much more electrical voltage than a pacemaker, with the capability of converting even a ventricular fibrillation to a normal rhythm. Studies show that implantable defibrillators do successfully convert ventricular fibrillation to correct arrhythmias [7].

In laboratory settings, insulin infusion pumps and glucose sensors under computer control can provide glucose control analogous to the human pancreas. Indeed, they are the components of a hoped-for artificial pancreas [8]. However, the development of automated systems for longterm, computer-controlled drug delivery has been stymied by the irascible behavior of biologic fluids. Clots and fibrosis confuse intravascular sensors and clog infusion pumps [9]. For now, the computer-controlled mechanical pancreas eludes us.

Open-Loop Control Systems

Unlike closed-loop systems, most medical computer feedback control systems have a human in the loop, who either gathers or sends the state variables, reviews or revises the computer's advice, or dispenses the medication.

Some of the oldest open-loop control systems are standalone computers that recommend dosages of digoxin, aminoglycosides, and theophylline. Typically, these programs estimate kinetic constants from the serum levels obtained after the first dose of the medication. These programs provide better drug levels than do unassisted physicians [10].

The computation of an optimal insulin dose is more difficult because insulin requirements depend on the patient's compliance with both dietary and exercise prescriptions and the timing of meals and exercise as well as the insulin dose and its excretion kinetics. Simple kinetic models do not provide perfect dose predictions. Furthermore, determining the "dose-response" of glucose control to exercise is expensive in both money and time, which is not practical in most settings [11]. Therefore, researchers have developed and tested more pragmatic models for both training and practice [12-16].

Home glucose monitoring systems can now retain the values of the glucose results and other variables for weeks to months, providing a convenient source of data for feedback control. Many vendors have tested products that collect patient data and transmit it to physicians by telephone for review. Some of these systems also provide advice about dosing. By alerting physicians early to trends, such systems offer the promise of improved glycemic control and better patient outcomes. In the largest and best randomized clinical trial of transmission of home glucose results to providers for review, Marrero and colleagues [17] found no difference in patient outcomes or levels of glucose control among intervention patients compared with control patients. However, the patients in the study already had good glucose level control at the outset, making it more difficult to show an effect. Smaller studies have shown some advantages to computer feedback [16], and noncomputerized suggestions to providers about heparin dosing did improve outcomes. Therefore, we are positive about the potential of computer-based insulin dose adjustment [18, 19].

How and when the feedback from computer protocols is delivered greatly influences the adoption of computer-based feedback systems. Stand-alone consulting systems that require the physician to go to a special device, initiate a dialogue, and describe the patient to the computer (for example, give the last blood urea nitrogen level, weight, insulin doses, and so forth) are not likely to be successful in a clinical setting, although they may be useful educational tools. Physicians will resist such systems because the benefits only rarely outweigh the time cost of using the system. Moreover, such systems cannot provide assistance unless the physician recognizes when he or she needs help and makes the effort to use the computer.

When the computer carries the relevant patient information in an electronic medical record system that the physician routinely accesses, it can offer reminders ("advice") without requiring extra physician effort or time, and physician acceptance comes more easily [20]. Almost all successful trials of open-loop feedback to physicians come from electronic medical record systems with automatic reminder capability [21-28], and such systems can have a large effect on the care process. Computer reminders produced increases up to fourfold in the use of preventive care among eligible patients in a long-term trial at our institution [2].

Open-loop rules can be quite complex. Figure 1 shows the rules that the HELP system at LDS Hospital in Salt Lake City, Utah, uses to govern the ventilator settings of patients with acute respiratory distress syndrome [22]. The computer generates reminders to physicians when the rules call for ventilator adjustments. Physicians, not the computer, then make the final decision about the ventilator settings.

View larger version (47K): [in this window] [in a new window]   Figure 1. Rules for acute respiratory distress syndrome. (Reprinted with permission from LDS Hospital.).

The protocols in open-loop systems can, on the other hand, be quite simple. Table 1 shows part of the rules that guide the use of influenza vaccine in the Regenstrief Medical Record System. The first rule in the Table definesexclusion criteria, for example, patients who have already either received or refused the vaccine during the past year. Succeeding rules suggest influenza vaccine for patients with chronic obstructive lung disease, or diabetes. Other rules (not shown) remind about using a flu shot in patients with congestive heart failure, renal failure, and advanced age (> 65 years old). The braces identify variables that will be replaced with patient-specific data when the reminders are displayed to the physician.

The second part of Table 1 shows one of many rules related to diabetes in pregnancy. The first block defines patients at high risk during pregnancy, and the second generates a message to the physician (or obstetrical nurse) about the needed test. The last example Table 2 shows yet another Regenstrief Medical Record System rule (in a somewhat different syntax) that reminds physicians to treat diabetic patients who have proteinuria with angiotensin-converting enzyme (ACE) inhibitors. Exclusion criteria are not shown in Table 2. The computer processes these rules and presents precomposed orders to a physician as she or he begins a computer order-writing session. The physician can accept or reject any suggested orders.

We have implemented many open-loop rules to guide the management of diabetic patients, including those to remind physicians to do the following:

- test glycosylated hemoglobin

- screen for cholesterol abnormalities

- avoid oral hypoglycemic agents in high-risk patients (for example, alcoholic persons)

- prescribe aspirin for diabetic patients who have at least one other cardiovascular risk factor

- treat hypertensive diabetic patients with ACE inhibitors to protect their kidneys

- refer patients for a fundoscopic examination at intervals, depending on the type and duration of diabetes.

Physicians wrote the rules shown in Tables 1 and 2 using a formal syntax that the computer can "understand" directly. Our institution does 70 000 medicine or medicine subspecialty visits, 30 000 obstetrical visits, and 30 000 pediatric visits each year, and before all of them, the computer reviews the patient's computer-stored record according to these rules and reminds the physician when it finds conditions needing attention.

The American Diabetes Association (ADA) regularly publishes several clinical practice recommendations for management issues [29]. Many of these could, in theory, also be implemented as open-loop computer reminders. However, some of these recommendations are difficult to transform into operational computer rules because they lack the precise criteria for specifying what action should be taken and when. In a patient with a foot ulcer, for example, the guidelines suggest that radiologic examination may be required to exclude subcutaneous gas, presence of a foreign body, osteomyelitis, and Charcot foot. This does not provide much direction. Is it really saying that one should obtain a radiologic examination at least once in most diabetic patients with foot ulcers? Or is it saying that one should obtain a radiologic study if one suspects a foreign body, osteomyelitis, or Charcot foot? The guideline goes on to say that a triple-phase bone scan, indium-tagged leukocyte study, magnetic resonance imaging scan, or bone biopsy might be required but gives no criteria for choosing among the four. The computer requires precise instructions for when to deliver a reminder.

On the other hand, the ADA guidelines provide some precise rules that would be easy to implement through a computer. For example, they suggest ordering tests for glycosylated hemoglobin, thyroid-stimulating hormone, thyroxine, high-density lipoprotein and low-density lipoprotein cholesterol, and triglyceride levels, urinalysis, and electrocardiography at the first visit of a diabetic patient more than 40 years old. Here, however, the scientific support for doing these studies is not included in the guideline report. Some authors suggest not routinely obtaining chest radiographs even for hospitalized patients in the absence of symptoms or special risks [30], and some would argue against routine electrocardiograms in the absence of cardiac symptoms or findings [31]. Most ADA guidelines about testing can be justified, but we wish that the scientific rationales were included in the report to save readers the requisite literature searches needed to support the guidelines locally. Such justification will be necessary if the ADA recommendations are to become accepted routine care; at our institution, these recommended first-visit tests cost $293.

Conclusions

In the United States, most diabetic care (85%) is provided by generalists rather than diabetologists [32, 33]. That ratio cannot change much, given the relatively small number of diabetologists. By producing and validating diabetes guidelines that could be implemented as open loops in electronic medical record systems, diabetologists could reach many more diabetic patients. First, however, more detailed justifications and more precise specifications are required. The need for precision in the guidelines is all the more important as we strive for tighter control of diabetic patients and skim closer to the hypoglycemic side of the treatment window. Eddy [34] and Tierney and colleagues [35] provide good advice about building guidelines, and Gunnar and colleagues [36] provide a great example of the value of tying guidelines to the supporting medical literature. However, information for deciding many of the details of diabetic care is likely not available in the current medical literature, and more primary studies are needed to define the precise relation between patient state variables and some of the actions needed. The Ottawa ankle guidelines provide a good example of how to do this [37].