Developing and Implementing Computerized Protocols for Standardization of Clinical Decisions

Humans have a limited ability to incorporate information in decision making (1, 2). Our short-term memory can simultaneously retainand therefore optimally utilizeonly four to seven data constructs; attempts to use larger amounts of information at one time lead to ineffective decision making (3). This limitation stands in striking contrast to the hundreds of variables encountered by clinicians in the clinical environment. The mismatch between human limitation and excess information almost certainly contributes to unnecessary variation in clinical practice (4-9), clinical error (10-22), and poor compliance with guidelines (23-27). In this paper, I consider the problems created by this mismatch and describe a computer-based decision-support system that successfully combines rules based on credible evidence with specific data from individual patients. I also review key social, psychological, and administrative elements that have made development of this system possible. Variation and the Standardization of Practice Variation in clinical practice persists even when guidelines based on reputable evidence are available (28, 29). Harm can result when clinicians do not comply with standard practice (9, 30, 31). Of note, much of current clinical practice has not yet been shown to produce more good than harm (18, 32-34). Widespread distribution of evidence-based guidelines (35, 36) and education programs (24, 37-40) have had only a limited effect on low clinician compliance, and patient compliance (41) and hospital compliance (42) are almost as low. Unaided human decision makers do not possess the consistency of behavior or the accuracy of perception necessary for the consistent delivery of recommended therapies (10, 43-48). Standardization of clinical decisions is needed not only for clinical practice but also for rigorous clinical research (49). Many interventions of clinical value have relatively small effects, with odds ratios of 3.0 or less (50). Systematically conducted clinical trials are necessary to recognize these small effects and to identify ineffective clinical care elements (50, 51). Without explicit methods, however, the fundamental scientific requirement of replicability of results (48, 49) cannot be achieved. An explicit method contains enough detail to generate specific instructions (patient-specific orders) without requiring judgments by a clinician. It is driven by patient data and generates the same instruction for a given set of input data. Any form of guideline or protocol can theoretically contain enough detail to constitute an explicit method. In practice, however, paper-based versions of any protocols but the simplest ones (for example, vaccination schedules or treatment of hypokalemia in a patient receiving digitalis and diuretics) cannot be made explicit and therefore remain dependent on clinician judgment. The need to standardize decisions provides a counterpoint to the equally compelling need to deliver individualized, patient-specific treatment. Unexpectedly, the discussed computerized protocols, which are explicit, detailed, and patient data-driven, can simultaneously achieve standardization of clinical decision making and individualization of patient therapy (52). Clinical care (the treatment a patient receives) is determined by clinical caregivers' decisions and by each patient's individualized expression of his or her illness (52). Consider an explicit method for mechanical ventilationfor example, a computerized protocolthat is used to standardize clinical decisions for two patients. The treatment given to one of the patients, who responded to positive end-expiratory pressure, would differ from that given to the other patient, who did not; treatments would differ even though the same explicit decision-making rules were used to standardize clinical decisions for both patients. The clinical care delivered to patients through computerized protocol instructions is therefore individualized (patient-specific) although decision making is standardized. Patient-specific care contrasts with time-driven decision-support tools, such as a clinical pathway that requires extubation within 36 hours and discharge from the hospital within a specified time. Unlike explicit methods driven by patient data (25, 53-58), time-driven tools raise legitimate concerns about patient-invariant (cookbook) care. One of the most attractive features of the use of point-of-care computerized protocols is their ability to individualize patient care while standardizing clinical decisions with an explicit method. The LDS Hospital Experience Two computerized protocols developed at LDS Hospital, Salt Lake City, Utah, form the case studies for this discussion. The first protocol standardizes bedside decisions for mechanical ventilation of patients with the acute respiratory distress syndrome (ARDS). These decision-support tools were initially developed for a randomized clinical trial of extracorporeal carbon dioxide removal in patients with ARDS (53). The tools were subsequently exported to 10 other hospitals (in 8 cities in 7 states) uninvolved in the development and were then evaluated in a randomized clinical trial of mechanical ventilation in patients with ARDS (54). The second protocol standardizes bedside clinical decisions for intravenous fluid and hemodynamic support. These protocols were recently developed for a projected randomized clinical trial of pulmonary artery catheters by the NIH/NHLBI (National Institutes of Health/National Heart, Lung, and Blood Institute) ARDS Clinical Network (55). Varieties of Decision Support Guidelines and protocols can effectively support clinical decision making (56) and can favorably influence clinician performance and patient outcome (28, 57-59). These decision-support tools have been functionally categorized as reminders, as consultants, or as educational (60). Thousands of decision-support tools with different names, focuses, and outputs are currently available to clinical practitioners. However, they often lack specific instructions for many of the scenarios encountered in clinical practice (Table 1). Most guidelines and algorithms (including guidelines generated by the Agency for Healthcare Research and Quality) are useful only in a conceptual sense (61-67); they neither standardize clinical decisions nor lead to uniform implementation of clinical interventions, although these are their ultimate goals (65, 67, 68). For example, it would be difficult to reduce variability with a protocol that required the clinician to determine whether the patient looked septic unless the state looked septic was explicitly defined by identifying the specific observations that lead to this clinical conclusion. When so defined, data that determine the presence of the state looked septic could be entered directly as inputs to a decision-support tool without the need for judgment by a clinician. Table 1. Decision-Support Products and Attributes Computerized protocols used for complex clinical problems can contain much more detail than textual guidelines or paper-based flow diagrams (67). Increased detail allows point-of-care generation of patient-specific therapy instructions that can be carried out by different clinicians with almost no interclinician variability (69). This can make both formal clinical inquiries (for example, randomized trials) and informal clinical inquiries (for example, some continuous quality improvement efforts or clinical practice evaluations) more robust (52, 69). Explicit Computerized Protocols Explicit rule-based computer systems are used in the most rigorous applications of algorithms in point-of-care clinical decision support (10, 25, 52-54, 67, 70-74). Computerized protocols have been shown to produce favorable changes in clinically important outcomes (10, 25, 27, 54, 57-59, 70, 72, 75-77). The immediacy of point-of-care decision support (67) differs from the more flexible time periods in which both decision trees (78) and computerized algorithms are implemented in consultative services (79). Although standardization of clinical decisions has been declared unreasonable (66), results from computerized protocol applications for mechanical ventilation suggest otherwise (53, 54, 73). Standardization of clinical decisions may have affected the conduct of certain randomized, controlled clinical trials (53, 54) in which success was previously thought to be unlikely (80). The computerized protocol strategy I describe falls within, or between, the categories of reminders and consultants outlined by Miller and Goodman (60). At LDS Hospital, computerized mechanical ventilation protocols have been used to support more than 250 patients with ARDS (Figure 1) for more than 100 000 hours (52, 53, 73, 85). Compliance of LDS Hospital physicians with protocol instructions was 95% (54, 86). These computerized ARDS mechanical ventilation protocols have been exported to stand-alone bedside personal computers in 10 other hospitals (Figure 2). These other hospitals have used the protocols to standardize clinical decision making for 103 patients with ARDS in a recently completed clinical trial (54, 87); physician compliance was 94%. A total of 38 546 instructions were generated during 32 055 hours of application (36% of instructions were for positive end-expiratory pressure, 40% were for fraction of inspiratory oxygen, 6% were for tidal volume, 12% were for ventilatory rate, and 4% were for minute ventilation) (54). Physicians objected to only 0.3% of the 38 546 instructions. Survival was comparable in the two experimental groups, but barotrauma was significantly less likely in the group that received standardized decisions from the computerized protocol. The 94% rate of physician compliance is much higher than the 30% to 60% rate seen in studies of computerized decision support protocols for antibiotic and diabetes guidelines (25, 27). These results indicate that an explicit method of care can