Chapter 13. Information Technology

The provision of healthcare is remarkably information intensive. A large integrated healthcare system processes twice as many computerized transactions each day (approximately 10 million) as the NASDAQ stock exchange (5 million) processes.

But high volume is just the beginning. Consider the task of tracking a single patient's current diseases, past medical history, medications, allergies, test results, risk factors, and personal preferences (such as for cardiopulmonary resuscitation). Tricky? Sure, but now do it over months or years, and then add in the fact that the patient is seen by many different providers, scattered across a region.

Want more? To make payment decisions, the insurer needs access to some of this information, as does the source of the insurance, which in the United States is often the patient's employer. But, because of privacy concerns, both should receive only essential information; to tell them of the patient's HIV status, or her psychiatric or sexual history, would be highly inappropriate, damaging, and possibly illegal.

Now let's make it really hard. Assume that the patient is in a car accident and taken to an emergency department (ED) in a nearby state, where she is stabilized and admitted to the hospital. Ideally, the doctors and nurses would see the relevant clinical details of her past history, preferably in a format that highlighted the information they needed without overwhelming them with extraneous data. Orders must be processed instantaneously (none of “the system is down for planned maintenance” or “orders are processed on the next business day” so familiar from commercial transactions). During the patient's hospital stay, not only would there be seamless links among all of the new observations (the neurosurgeon can easily view the ED doctor's notes; the resident can quickly find the patient's vital signs and laboratory studies), but also the various components would weave together seamlessly. For example:

• The system would prompt the doctor with information regarding the appropriate therapy or test for a given condition (along with links to the evidence supporting any recommendations).
• The system would warn the nurse that the patient is allergic to a medicine before she administers it.
• The system would tell the doctor or pharmacist which medications are on the formulary and steer them to the preferred ones.

Meanwhile, the vast trove of data being created through this patient's case—and millions of others like it—would be chronicled and analyzed (“mined”), searching for new patterns of disease, evidence of preferred strategies, and more. All of this would be iterative—as new information emerged from this and other research about disease risk factors or best practices, it would seamlessly flow into the system, spurring the next patient's care to be even better.

Contrast this vision of information nirvana with the prevailing state of most doctors' offices and hospitals.1,2 Information is stored on paper charts, and thus unavailable to anyone who lacks physical possession of the relevant piece of paper (in some cases, the notes are sufficiently illegible that even physical custody of the paper does not insure information access). Notes are entered as free text, not in a format that facilitates analysis or productive interaction with other pieces of system data. When the patient moves across silos—from outpatient to inpatient, from state to state, from hospital to hospice—crucial information rarely moves with her. Communication of facts (e.g., medication lists, allergies, past medical history), which should be streamed through the system, instead lives at the mercy of person-to-person interactions or a haphazard pinball game of photocopies bouncing from place to place.

Even at the level of the individual practitioner, the impact of this chaos is profoundly demoralizing, wasteful, and dangerous. Just watch a nurse take a patient's vital signs on a typical hospital ward. The nurse looks at the numbers on the screen of a digital automated blood pressure cuff: 165/92. She records them on an index card (or, sometimes, on her forearm or the cuff of her scrub suit), hopefully next to the correct patient's name. Later, she returns to the nurses' station and transcribes these numbers (again, hopefully belonging to the correct patient) onto the appropriate place in the chart (hopefully the right chart). Then, in a teaching hospital, an intern transcribes these vital signs onto another index card (or maybe, these days, an iPad) during morning rounds. He presents these data to his resident and later to his attending, who each do the same thing. Eventually, each of these practitioners writes, or dictates, notes for the medical record. Any wonder that this information (which, you'll recall, began life in digital form!) is frequently wrong? Or that busy healthcare professionals find that huge chunks of their valuable time are squandered? Or that the patient has the sense (particularly after she has been asked the same question by 10 different people) that the right hand has no idea what the left hand is doing?

Why has healthcare, the most information intensive of industries, entered the modern age of computers so sluggishly, reluctantly, and haphazardly? Part of the reason is that, until recently, the business or clinical case for healthcare information technology (HIT) was far from ironclad. Such justification was needed because HIT is extraordinarily expensive (about $50,000 per doctor in an ambulatory office, and up to $150 million to completely wire a large teaching hospital), unreimbursed, and extremely challenging to implement. Moreover, until the past few years, most healthcare computer systems were relatively clunky and user unfriendly, in part because the market for them was too weak to fund the research and development and generate the user feedback and refinement cycles needed for complex systems to mature.

We are now at a turning point for HIT. The patient safety movement has catalyzed the widespread recognition that a robust computer scaffolding is absolutely essential to efforts to improve quality, safety, efficiency, and coordination of care. There are striking examples of successes, such as that enjoyed by the huge Veterans Affairs (VA) system, whose early adoption of electronic health records (EHRs) and computerized provider order entry (CPOE) sparked a substantially improved quality of care.3 A large and convincing literature demonstrates that well-designed and implemented electronic systems can lead to significant benefits for systems and patients.4–8

At the same time, there is a more troubling, though not entirely surprising, side of the story. A decade ago, my colleagues and I openly worried that the early glowing evidence about HIT's benefits had come from a handful of institutions that had lovingly built their own systems, had highly committed leaders, and had invested heavily in computerization.9 As Shojania and I wrote in our 2004 book Internal Bleeding:

But the average hospital will not share these conditions, any more than your local Gilbert and Sullivan troupe resembles the Metropolitan Opera … More than one CIO has tried to airlift a commercial system into her hospital, then stood scratching her head at how slick the system seemed to be during the vendor's demo, and how poorly it performed in real life.10

As you've already seen, our suspicions—that the implementation of commercial IT systems at thousands of hospitals and hundreds of thousands of ambulatory offices might not go quite so smoothly—have, sadly, been proven correct. A growing literature tells stories of HIT implementation failures, unanticipated consequences, and even harm.11–16 It is a sobering but useful tale, one that highlights the immense challenges of fundamental change in complex adaptive systems (Chapter 2), the critical need to use human factors design principles (Chapter 7), and the overwhelming complexity of the task itself.

This chapter will describe the main types of HIT systems, some of their safety advantages (Table 13-1), and some of the problems—including new kinds of errors—they can create. Because HIT addresses so many safety targets and is a component of so many solutions, considerable information about HIT's role in specific areas can also be found throughout the book.

Because most medical errors represent failures in communication and data transmission (Figure 9-3), computerization of the medical record would seem like a safety lynchpin. (While the old term was “electronic medical record” [EMR], the term “EHR” is generally preferred, since it emphasizes the role of the patient in viewing and even contributing to the record and the fact that the EHR may chronicle and even influence health status, not just a patient's medications and diagnoses.) But to realize this benefit, attention must be paid to a variety of system and user factors. The system factors include the ease of use, the speed with which data can be entered and retrieved, the quality of the user interface, and the presence of value-added features such as order entry, decision support, sign-out and scheduling systems, links to all the necessary data (e.g., x-rays and electrocardiograms), and automatic reports. The user factors primarily relate to the training and readiness of the provider and nonprovider workforce (Chapters 7 and 16).

User efficiency is particularly important. Despite the hope that computerization would save time for providers, emerging evidence indicates that the opposite is often true, particularly for physicians.17 Some of this cost in time may be repaid in more efficient information retrieval, but increasing attention will need to be focused on facilitating workflow. (Remember that digital blood pressure reading?—in the “wired” hospital, it will magically leap from blood pressure machine into the EHR, where it can be seamlessly imported into each provider's note.) Effective systems will, of course, provide huge efficiency benefits to administrators, researchers, and insurers by capturing data in standardized formats and allowing electronic transmission.

Unfortunately, this facilitated movement of bits and bytes has a dark side, in the form of the “copy-and-paste” phenomenon. One tongue-in-cheek essayist captured the problem beautifully:

The copy-and-paste command allows one day's note to be copied and used as a template for the next day's note. Ideally, old information and diagnostic impressions are deleted and new ones added. In reality, however, there is no deletion, only addition. Daily progress notes become progressively longer and contain senescent information. The admitting diagnostic impression, long since discarded, is dutifully noted day after day. Last month's echocardiogram report takes up permanent residence in the daily results section. Complicated patients are on “post-op day 2” for weeks. One wonders how utilization review interprets such statements.18

A study of 167,000 computerized records in the U.S. Department of Veterans Affairs (VA) healthcare system found that physical examinations were completely copied (from one author to another) in 3% of the charts.19 Some IT systems can disable the copy-and-paste function, although this may be unacceptable to providers (some information really does remain static from day to day, and having to retype or redictate it is wasteful and annoying), and resourceful providers can usually find a way to bypass such restrictions anyway.20 In the end, copy-and-paste is like many elements of patient safety, in that introducing structural changes without ensuring that caregivers possess the requisite education and professionalism is likely to lead to new types of mischief, or even harm.

One important choice for EHR developers is the use of structured versus unstructured data. Physicians, in particular, have always preferred entering data as unstructured prose (captured through typing, transcription services, or increasingly, speech recognition software), which has the advantages of ease of use and familiarity, and is consistent with the long tradition of “telling the patient's story.”21,22 On the other hand, unstructured data are difficult to analyze for patterns, to use to support quality improvement and public reporting activities, and to link to computerized decision support. Structured, or coded, data (captured via templates or automatically fed into the EHR after being recorded by electronic devices) can support these functions, but they fail to comport with clinicians' mental model for data acquisition and analysis. Ongoing research is attempting to identify the best mix of structured and unstructured EHR inputs, and to develop new ways (through “natural language processing”) to mine narrative prose for key data elements.22 It is likely that high-quality EHRs of the future will combine the best of both types of data capture and analysis.

An important advance in EHRs is the ability to promote patient engagement (Chapter 21). At the very least, allowing patients access to their laboratory data and giving them the ability to schedule their own appointments represents real progress.23 But patients, who are increasingly accustomed to managing their own affairs (financial, travel, dating) with the support of electronic systems, are interested in new kinds of tools to help with their medical care and overall health.24 Many patients now keep personal health records (PHRs), although a minority do so electronically. Future developments in this area are likely to weave patient- and provider-facing electronic systems together, facilitate new kinds of information flow and communications (e-mails, text messaging, video consultations, and more), and raise a host of issues surrounding reimbursement, privacy, data integrity (what happens when the provider and patient disagree about what should be in a jointly created record?), and more.25,26 Obviously, the implications of this on patient safety will be profound, with the capacity both for great leaps forward and some new hazards.

EHRs raise yet another challenge, one that is even more profound. As the physician-writer Abraham Verghese has observed, clinicians' focus is increasingly centered on the data in the computer, sometimes at the cost of the human connections so fundamental to the practice of medicine and the art of healing. After discussing the traditional approach to patients, in which “the [patient's] body is the text,” Verghese writes of a new, more expedient approach that he sees in today's trainees:

The patient is still at the center, but more as an icon for another entity clothed in binary garments: the “iPatient.” Often, emergency room personnel have already scanned, tested, and diagnosed, so that interns meet a fully formed iPatient long before seeing the real patient. The iPatient's blood counts and emanations are tracked and trended like a Dow Jones Index, and pop-up flags remind caregivers to feed or bleed. iPatients are handily discussed (or “card-flipped”) in the bunker [the conference room where the team does its work], while the real patients keep the beds warm and ensure that the folders bearing their names stay alive on the computer.27

Taking full advantage of the remarkable potential for EHRs to improve safety and quality while heeding Verghese's cautionary tale may be one of the most subtle yet important challenges facing caregivers in the modern age.

Because the prescribing process is one of the Achilles' heels of medication safety (Figure 4-3), efforts to computerize this process have long been a focus of safety efforts. In 1998, Bates et al. demonstrated that a CPOE system with decision support reduced serious medication errors by 55%, mediated by improved communication, better availability of information, constraints to prevent the use of inappropriate drugs, doses, and frequencies, and assistance with monitoring.4,28 Another study of more sophisticated decision support found an 83% reduction in medication errors.29 The advantages of CPOE over paper-based systems are many; in addition to those listed in Table 13-2, the installation of CPOE systems inevitably leads organizations to standardize chaotic processes (the equivalent of cleaning out your closet before moving), which has its own safety benefits.4

Much of the value of CPOE systems comes from identifying out-of-range results or potentially unsafe interactions, and rapidly alerting providers so that they can decide whether their plan is correct.30 For example, a CPOE system can alert a provider to a potentially fatal medication–allergy interaction (Figure 13-1) or a potentially dangerous laboratory result (Table 13-3). These systems can also be used at the healthcare system level to identify and track errors via trigger tools (Chapters 2 and 14).

In addition to helping clinicians avoid mistakes, CPOE systems can suggest actions that should always accompany certain orders. These “corollary orders” should be second nature, but our memories are fallible and we sometimes forget to check a creatinine and potassium after starting an angiotensin-converting enzyme (ACE) inhibitor, a digoxin level after beginning digoxin, or a glucose level after starting insulin. In one study, clinicians were twice as likely to accept a computerized prompt for a corollary order than they were to place the order without the prompting.31 I'll say more about these functions—which go under the general name of clinical decision support systems (CDSS)—later in this chapter.

Despite the great appeal of CPOE, one study at an academic medical center found 22 new error types caused or exacerbated by a commercial CPOE system, including long gaps in medication delivery because of a fragmented CPOE display, failure to discontinue medications or renew antibiotics, and delayed ordering caused by CPOE system downtime.11 Even more worrisome, one prominent children's hospital experienced a threefold increase in the mortality rate of critically ill patients after a new CPOE system was installed.12 There, caregivers noted inefficient order entry, too much time spent at the computer screen and away from the bedside, and several other workflow-related problems. While this study has been criticized on methodological grounds and a subsequent study of the same CPOE system at another children's hospital found improved mortality,8 the cautionary notes it raises about human factors (Chapter 7) and unanticipated consequences merit our attention.

We can expect that, as the market for CPOE grows and commercial products experience many crucial user-feedback-generated improvement cycles, the systems will become better, errors associated with them will become less common, and the full safety benefits of CPOE will begin to be realized. Table 13-4 highlights the many categories of IT-related errors that have been identified and suggests a number of mitigating strategies.15

Bar Coding and Radio-Frequency Identification Systems

Even when rigorous safety checks are embedded into the prescribing process, errors at the time of drug administration can still lead to great harm (Chapter 4). To prevent these errors, many institutions are implementing bar coding or radio-frequency identification (RFID) solutions. In bar code medication administration (BCMA), a nurse must swipe a bar code on the medication, the patient's wristband, and her own badge to confirm a three-way match before a medication can be administered.32 In RFID systems, the medication package has an implanted chip that transmits a signal, allowing for passive identification (like driving through an automated toll booth) rather than requiring a scan. Despite its intuitive appeal, RFID remains more expensive, and—because patients are taking multiple medications and nurses often have the medications for multiple patients on their carts—somewhat trickier to implement. For now, most hospitals seeking to improve their medication administration process have favored BCMA.

Like all HIT systems, BCMA has its challenges. Nurses worry that it will take too much time (although one study demonstrated that it did not33)—and some observers have already documented workarounds (such as when a nurse takes a handful of patient wristbands and scans them outside the patient's room to save time) that bypass the systems' safety features.34 Another concern is whether BCMA systems can be sufficiently flexible to deal with acutely ill patients.35

Finally, BCMA must be rooted in an environment of robust safety processes. We documented one case in which two patients (one a poorly controlled diabetic) were mistakenly given each other's bar-coded wristbands, nearly leading to a fatal insulin overdose in the nondiabetic whose computerized record erroneously indicated that he had a stratospheric blood sugar.14 Like all IT systems, BCMA can become a very efficient error propagator if the inputted data are incorrect.

These concerns notwithstanding, effective use of BCMA technology can substantially reduce medication dispensing errors. After a long period during which bar coding was supported more by face validity than hard evidence, two studies by Poon et al. have demonstrated that bar coding really does work. The first focused on a process invisible to most doctors and nurses: the drug dispensing system in the pharmacy. There, a bar code scanning system reduced both errors and potential adverse drug events.36 The other study, of a bedside bar coding system, is probably the more important one for the field of patient safety.7 In it, a “closed-loop” system that combined CPOE, BCMA, and an electronic medication administration record (eMAR) led to an approximately 50% decrease in drug administration errors and potential adverse drug events.

I have come to believe that bar coding may be even more important than CPOE for hospital medication safety, at least in the short term. In the last several years, there have been a number of high-profile medication errors in the United States, virtually all involving medication administration (including the heparin dosing error that nearly killed actor Dennis Quaid's newborn twins37). In these cases, the problem was that there was nothing standing between a nurse and a potentially fatal mistake: no layers of metaphorical “Swiss cheese” (Chapter 2). During this same period, I cannot recall a single prescribing error that garnered similar attention.

Does this mean that prescribing errors are a thing of the past? Sadly, of course not. But even in the many hospitals and offices that lack CPOE, many prescribing errors are now avoided because physicians have access to handheld or computerized prescribing aids, allowing them to quickly look up drug indications, interactions, and dose ranges. Moreover, some of the riskiest practices—such as writing “U” for “units”—are now banned by the Joint Commission (Chapter 4).

Whether my admittedly unscientific observation is true or not, it is comforting that BCMA, which seemed like a good idea for such a long time, is now supported by robust evidence. In the end, of course, hospitals will need both BCMA and CPOE. The studies of Poon et al.7,36 demonstrate the synergistic value of blending these technologies together in a “closed loop.”

Smart Intravenous Pumps

Progress in medication safety through BCMA still leaves a large gap: the safety of medications infused intravenously. Approximately 90% of hospitalized patients receive at least one intravenous medication. Because many of these medications are far more dangerous than pills, and their doses are more variable (often calculated by the hour, or through complex weight- and size-based formulae), the opportunity for harm is real. Here again, as with BCMA, the problem is that there is no downstream opportunity to catch errors at the administration phase. Because of this, there has been considerable interest in so-called “smart intravenous pumps.”

Smart pumps are engineered to have built-in danger alerts, clinical calculators, and drug libraries that include information on the standard concentrations of frequently used drugs. They also can record every infusion, creating a database that can identify risky situations and medications for future interventions. Studies have shown that these pumps can prevent many infusion errors,38–40 but that attention must be paid to seamlessly interfacing these systems with other computerized medication systems such as CPOE and BCMA.41

One group of researchers compared three types of pumps—traditional pumps, smart pumps, and smart pumps with bar coding technology—and found that the smart pumps were most effective either when they had bar coding (to ensure that the pharmacy-prepared bag was the correct one for a given patient) or when “hard” (i.e., unchangeable by the bedside nurse) dosing limits were used.42 When one considers the increasing number and complexity of intravenous infusions in hospitalized (Figure 7-1), and now even homebound, patients, perfecting this technology and working through the complex machine–person interaction issues should be high priorities.

Other IT Solutions

When people think of information technology and patient safety, they generally think of EHRs, CPOE, and perhaps BCMA. It is worth adding several others to the list of IT tools that can help improve medication safety, such as “smart pumps,” eMARs, and automated drug dispensing systems.43 Once the bugs are worked out of these systems, hospitals' challenges will be to find the money to purchase, maintain, and train people on all of them, and then to weave them together into a seamless whole (Chapter 7).

It is worth pointing out that a wide range of other IT-based solutions can help improve patient safety outside the sphere of medication safety. For example, in many hospitals, staff members now wear voice-activated wireless microphones, or use modern text paging or cell phone systems—even systems modeled on Facebook and Twitter—to facilitate instant communication among caregivers. The value of IT-based sign-out systems and computer-based simulation is discussed in Chapters 8 and 17, respectively. And we shouldn't forget the importance of more clinically oriented HIT, such as the picture archiving and communication systems (PACS) that allow digital radiographs to be reviewed from a few miles, or a few thousand miles, away from the hospital or clinic.44 In addition to their convenience, PACS can decrease x-ray interpretation errors by facilitating double reads, computerized enhancements of images, and access to prior radiographs.45 Moving even closer to the patient, the use of handheld ultrasound can lower the risks of central line placement or thoracentesis (Chapter 5).46,47

Although much of HIT's emphasis has been on replacing the paper chart and moving information around, its ultimate value may lie more in computerized clinical decision support systems (CDSS) that can provide guidance to clinicians at the point of care. For example, some systems offer simple alerts such as drug–drug, drug–allergy, or drug–lab interactions (Figure 13-1),48 or links to evidence-based guidelines (the clinician types in a diagnosis of “pneumonia” and a link to a recent pneumonia management guideline materializes).

But that is just the start. More prescriptive decision support systems can “hard wire” certain kinds of care. For example, order sets for common diagnoses can be loaded into a CPOE system, making it easy to do the right thing by simply clicking a few boxes.6,49 Or an intensive care unit (ICU) system can alert the physician or nurse when a patient's vital signs veer outside preset parameters (Figure 13-2). Note that these prescriptive systems usually permit clinicians to deviate from recommended protocols, but this takes more time, because the doctor needs to type out the orders instead of accepting an order set, and may even be asked to document the reason for deviation. Even more prescriptively, the computer could all but force a given practice, making the clinician jump through several hoops (such as “call a specialist for approval”) before being allowed to bypass a recommendation.

Of course, this is tricky stuff, as humans worry both about machines taking over their lives and that guidelines embedded in computer systems may not completely apply to the situation and patient at hand. Many healthcare experts and providers are also troubled by the possibility that prescriptive guidelines may stand in the way of innovation or “out of the box thinking.” These issues are real and will need to be worked through over time, but for healthcare systems, which are increasingly being judged on their adherence to evidence-based processes of care (Chapters 3 and 19), the level of tolerance for individual practice differences that vary from established best practices is likely to fall.

More pragmatically, one of the great challenges of CDSS systems has been alert fatigue, as clinicians rapidly tire of being bombarded by warnings and sometimes fail to notice even the important ones.50 One study of approximately 5000 computerized alerts showed that clinicians overrode “critical drug interaction” and “allergy–drug interaction” alerts approximately 75% of the time.51 And these alerts may anger clinicians, particularly if they are time-consuming to address. In one famous case, an expensive CPOE system failed partly because physicians rebelled against all the alerts.13 As systems become more prescriptive, clinicians may also bristle at hardwired care protocols that appear to lack flexibility (“cookbook medicine”).

All of this creates a daunting calibration challenge for designers of CDSS: making sure clinicians are alerted when appropriate but not so much that they simply learn to ignore or click out of every alert (especially the important ones). A recent JAMA commentary even raised the specter of physicians or health systems forestalling their purchase of CDSS systems because they fear being held liable for ignoring alerts (Chapter 18).52 Because the experience of virtually every HIT implementation has been that the vast majority of alerts are ignored, most modern installations begin with a “less is more” philosophy: retaining the truly life-threatening, mission-critical alerts while jettisoning the “nice to know” ones. This philosophy, which has been taken to high art by Google (think about how uncluttered your Google search page is) and Apple (ditto your iPad), is deeply rooted in human factors thinking and also has implications for alarm systems in ICUs and hospital wards (Chapter 7).

Along with our growing awareness of the human factors implications of alerts, we are also starting to see the blending together of computerized systems and multidisciplinary patient safety and quality improvement programs. In Chapter 11, I discussed a fall prevention protocol that was embedded in a computerized environment: the IT system calculated the patient's risk of falling and created a customized educational program for the patient and family, while also targeting actions for the various providers (Figure 11-2).53 We can expect to see more of this type of integration—an exciting development.

Just as exciting, a new philosophy of “measure-vention” is emerging, in which real-time computerized measurement of performance is not fed back to providers one or two weeks (or one or two years) later, but immediately, allowing for “just in time” improvements. For example, one program aims to improve adherence with venous thromboembolism prophylaxis (Chapter 11) by using an electronic dashboard that captures and reports each patient's VTE prophylaxis status in real time (Figure 13-3).54 In another application of the same principle, a real-time electronic surveillance dashboard allows pharmacists to intercept and prevent medication errors and optimize therapy.55 These innovative strategies hold real promise for improving safety.

Finally, another type of decision support focuses on improving diagnostic accuracy (Chapter 6). Early clinical artificial intelligence programs—in which clinicians entered key elements from the history, physical examination, and laboratory studies and the computer fashioned a list of diagnosis—were disappointing, because the computer-generated diagnostic lists mixed plausible possibilities with nonsensical ones, and the data entry time (which came over and above clinical charting time) was prohibitive.56

Recent advances have generated new interest in diagnostic decision support. Some programs now pull data directly from the EHR, bypassing the need for redundant data entry. Others mine textbooks and journal articles to find diagnoses most frequently associated with citations of certain symptoms and signs.57 Most modern programs not only suggest possible diagnoses but also link to helpful resources and references. The accuracy of today's top programs, such as Isabel (Figure 13-4) and DxPlain, far surpasses that of earlier models.58,59

Diagnostic decision support systems are likely to be judged on the seamlessness of the inputting process and the helpfulness of the diagnostic possibilities they offer. As of yet, no studies convincingly prove that these systems improve patient outcomes, but this is probably more reflective of the challenges of measuring the frequency and impact of diagnostic errors (Chapter 6) than it is of the quality or usefulness of the tools.

One can envision future computerized decision aids that draw their information directly from the EHR (though this will be challenging if patient histories remain mostly unstructured, as I discussed earlier in the chapter), produce possible diagnoses that are automatically updated with new information, and actually “learn” by integrating prior experiences from the system itself, making them ever more accurate over time. In a world in which http://www.Amazon.com helpfully suggests, “People who bought book X also liked book Y,” how about a decision support system that says, “Patients at our hospital with this constellation of findings often turned out to have diagnoses X or Y”? While this might have seemed like fantasy a decade ago, anyone who observed the IBM computer “Watson” defeat the world's top Jeopardy! contestants became convinced that it may be just a matter of time before this kind of computerized firepower transforms the diagnostic process. Oh, and by the way, IBM is working on a medical version of its Jeopardy!-playing computer: “Dr. Watson.”60

The story of HIT neatly follows along the lines of what has been called the “Technology Hype Cycle” (Figure 13-5).61 The Cycle, originally described by the consulting firm Gartner, is made up of a predictable series of phases that technologies tend to traverse after their introduction:

• “Technology Trigger”: after its initial launch, the technology reaches the attention of the public and industry.
• “Peak of Inflated Expectations”: A few successful applications of the technology (often by highly selected individuals or organizations—sound familiar?) catalyze unrealistic expectations, often pushed along by word of mouth, the blogosphere, or the marketing prowess of the vendor.
• “Trough of Disillusionment”: Virtually no technology can live up to its initial hype. As negative experience grows, the balloon is pricked and air rushes out. The media move on to cover another “hotter” technology, and the cycle repeats.
• “Slope of Enlightenment”: A few hardy individuals and organizations, seeing the technology's true potential, begin experimenting with it, no longer burdened by inflated expectations. Assuming that the technology is worthwhile, they begin to see and demonstrate its value.
• “Plateau of Productivity”: As more organizations ascend the “Slope of Enlightenment,” the benefits of the technology (which has improved from its initial clunky phase based on user feedback and the self- preserving instincts of the developers) become widely demonstrated and accepted. The height of the plateau, of course, depends on the quality of the technology and the size of its market.

This cycle perfectly describes the history of CPOE (over the last 15 years) and diagnostic artificial intelligence tools (over the past 30–40 years). The hope is that these technologies—all HIT, really—are now somewhere on either the Slope of Enlightenment or the Plateau of Productivity.

In any case, while I believe that we are now more realistic about HIT's warts (and that's a good thing), there is no doubt that we do need to wire our healthcare system, that some speed bumps are inevitable, and that the HIT story represents a market failure. No federal intervention was needed to get United Airlines or Wal-Mart to computerize—the business case to do so was compelling. But in healthcare, the fact that HIT systems are so complex and expensive, that many of the benefits accrue to parties other than those shelling out the money (such as insurers), and that HIT systems need to “talk to each other” (interoperability) creates a compelling case for some degree of government involvement.62

In the United States, the federal government is in the process (in 2012) of distributing approximately $30 billion to hospitals and community physicians who implement HIT systems that meet a variety of functionality and interoperability standards known as “Meaningful Use” (Table 13-5).63,64 The presence of such support in nearly all developed countries (more directly in countries with government-run healthcare systems) signals the start of a very rapid implementation phase for HIT worldwide. Implementation is also being catalyzed by studies illustrating a positive business case65,66 and by pressure from patient safety advocacy groups, such as the Leapfrog Group, a coalition of major business purchasers that has deemed HIT to be one of the “evidence-based safety practices” that it uses to judge hospitals.67

As self-help guru Tony Robbins once observed, most people overestimate what they can do in a year, and underestimate what they can do in a decade. In the case of HIT, this is likely to be an accurate assessment. While adoption will continue to occur in fits and starts for the next few years, a decade from now HIT systems will have transformed much of the safety enterprise, if not all of healthcare. If the systems mature and developers and implementers learn from past mistakes, this will be decidedly for the better.

• The implementation of HIT has been remarkably slow until recently, but the pace is beginning to accelerate, driven by a growing business case and government incentives.
• Many healthcare activities require multiple providers (and others) to view patient-level information simultaneously, a powerful argument for EHRs.
• CPOE can ensure that physicians' orders are legible and respect preset parameters.
• Bar coding or other similar systems can help decrease the frequency of medication administration (and other patient identification–related) errors.
• Ultimately, much of the benefit of HIT will come through the thoughtful implementation of computerized decision support, which ranges from simply providing information to clinicians at the point of care to more prescriptive systems that “hardwire” certain care processes.
• While HIT has tremendous appeal, its history to date has been characterized by a surprising number of failures and unanticipated consequences, illustrating the importance of human factors and the challenges of implementing major changes in complex organizations.