On Informatics

The focus of my research is clinical decision support. Decision support means different things to different people. In my case it means I work on tools (usually computer tools) designed to help clinicians make better decisions.

There’s powerful evidence that decision support can yield better care, improve outcomes and save lives – but only where it’s used. The most advanced decision support systems are in use at a handful of academic medical centers such as LDS Hospital in Utah, the Regenstrief Institute in Indiana, Vanderbilt and the Harvard-affiliated hospitals in Boston. These centers have seen great benefits from their decision support, but decision support systems have been much less widely used in non-academic community settings where the majority of healthcare is delivered. Researchers have speculated widely on the cause for this disparity, but most explanations center on resources – academic medical centers often have large committees and researchers who can devote time and effort to developing good clinical decision support, and they’re much more likely to have developed their own clinical information systems. Community settings, by contrast, are much less likely to have these luxuries. There needs to be a way to get decision support from the “have” to the “have-nots”.

I recently had the opportunity to work with a national panel of decision support experts to consider this problem of transferability, and other pressing issues in clinical decision support. Our work was commissioned by the Office of the National Coordinator, and culminated in a recently released Roadmap for National Action on Clinical Decision Support. This report focused on actionable steps the healthcare community can take to advance decision support efforts, such as:

• Starting pilot projects in community care settings to find out what’s needed to enable these settings to successfully adopt and use clinical decision support.
• Assembling a starter set of decision support content that can be readily adopted by community hospitals and care providers.
• Improving standards for knowledge representation, modelling and sharing.
• Conducting fundamental research into technical methods and architectures for decision support, as well as into organizational issues that affect decision support adoption.

The long term goal, of course, is uniform adoption of good clinical decision support in all care settings. We’re not there yet, but recently, we’ve seen rising adoption rates for clinical systems, and an increased national focus on IT – both key enablers for clinical decision support. There’s still a long way to go before the decision support “have-nots” become “haves”, but I’m optimistic that we can get there, and excited to see what the road ahead will bring.

NY Times article about Microsoft’s leap into health care. The company bought the Azyxxi software which was developed at the Washington Health Center (up the road) and has been implemented at a half-dozen other hospitals. The software is a cross-platform “data exploration engine.”
Article:

http://www.nytimes.com/2006/07/26/technology/26cnd-soft.html

More on Azyxxi at:

http://www.imedi.org/dataman.pl?c=lib&dir=docs/Azyxxi

The pre-publication draft of the IOM report “Quality Chasm part IV: A New Hope”, er, I mean, “Medication Errors” made its PR round last week. What does this mean for informatics?

First, let us set the stage:

The study says that every patient per day of hospitalization, on average, will have a medication error (even leaving out the less interesting ‘wrong time’ errors). If correct (see below), this means there are about 175 million errors per year in hospitals (38.2 million hospitalizations with 4.6 days mean LOS).
Then, they calculate actual injuries – adverse drug events – from several studies as 1.5 million, costing $3.5 billion per year. This includes inpatient and outpatient injuries (actually higher number for outpatients and long term care).

Finally, they make a series of recommendations (all heavily summarized):

1. Inform consumers (and give them the right to ask and refuse);
2. Improve resources for consumers (led by the National Library of Medicine);
3. Give providers the knowledge and tools they need to maximize safety;
4. FDA should improve labeling, naming, and nomenclature of medications;
5. Do implementation and structural research (NLM and AHRQ);
6. Congress should fund 1-5; and
7. We should explore policies to incentivize best practices (including pay for medication performance).

When do we care about medication errors?

We care about them because some lead to injuries and death. This is by far the most important concept I want to get across, and its obvious converse – the ones that don’t lead to injuries and death (and had little chance to) are far less important from a societal sense. (yes, from a structural and risk sense, but if we are holding health delivery accountable to these numbers, we must focus!)

We also care about the ones that are fixable, e.g. preventable or ameliorable. Preventable is inevitably a judgment call, although the calls frequently made have a lot of face validity. If we know of a systematic way to reduce these kinds of errors, then the error is preventable. If not, we should look for ways. Ameliorable really relates to the amount of injury that is caused by the error. If, for instance, as happened at Dana Farber more than a decade ago, a massive, deadly dose of chemotherapy is prescribed in error. If, then, in our imaginary world (it led to the death of a patient in the instance given), it is to be given over a period of 3 hours, and started but not completed because of some fail-safe system, then the injury may be far less than otherwise. So, systems to lessen harm are very important, especially where not preventable.
Are these numbers right?

The report freely admits that it is not a systematic review, nor are the data good for many areas. The numbers, therefore, are rough. Their primary solution is to underestimate the numbers as much as possible; for example, if you looked at various related studies, you might get 7-12 million ADEs in hospitalizations (7% of the errors in hospitals related to prescribing and administration).(1) If you looked at a review of ADRs, you might see $75 billion in costs versus $3.5 billion.(2) They do appropriately exclude the ‘wrong time’ errors, but they neglect seriously to examine the major study designs used. The techniques of identification have varied widely over the years; the best recent paper looking at this issue is by Nebeker.(3) With varying techniques, their approach to simply add together the numbers from a set of core papers is fraught with peril; one neither knows whether the ADEs are redundant nor completely disparate (apples and oranges). Nonetheless, dangerous medication errors exist and cause harm.
Does this paper help us move forward towards the available solutions?

All of the recommendations seem reasonable. However, the core benefits are hidden in a large amount of text which is not weighted. There are lists for individual groups to start doing – patients, providers, pharmacists – which are dozens of items long. The seven recommendations do not tie what might be most successful with the recommendations. If I were to shorten the list, I would put the following:

1. Patients assertively ask for best practices. I completely agree with the concept, but achieving this is a challenge. One part that I think is necessary is easier methods for having the patient and family update their own medication list as they transition across care elements.
2. Electronic prescribing with decision support. Not sufficient because needs to be done carefully and studied to understand consequences.
3. Health Information exchange to move towards complete and accurate medication lists (has to get used; implies standards).

Recently, I (re)discovered this gem of an article:

	Powsner SM, Tufte ER.	Related Articles, Links
	Graphical summary of patient status. Lancet. 1994 Aug 6;344(8919):386-9. No abstract available. PMID: 7914312 [PubMed - indexed for MEDLINE]

One enormous difficulty in the care of complex patients – in this case, hospitalized patients – is the sheer volume of information that is generated over relatively short amounts of time. One of the key requests seen by EHR developers in these settings (whether ‘home-grown’ or vended) is to better summarize the data. This has led to many different solutions in the area, from flexible flowsheets (think excel worksheets) to ’rounds reports’ (single page summaries of key information usable as the health care team goes on rounds) to snapshots (special pages in the EHR that attempt to summarize various aspects of health or care).

For those familiar with the use of EHRs, all of the above examples likely ring a bell. In my experience, these components are often the value-added aspects of the EHR for resident physicians and students in academic centers – the ability to have, easily at hand, a wealth of knowledge about patient status. Interestingly (and, I admit, anecdotally), these summaries are often printed. We created such a summary sheet (available for view here) and found use was higher when printed and placed on the door prior to a visit. The physician or nurse, typing madly into the EHR, would refer to various aspects when making orders or creating notes, alleviating the need to ask the system to flip back and forth between old data and new entries. The sheet provided a lot of pertinent information about specific aspects of a patient’s health in a simple format. The major complaint was still the amount of data one needed to sift through.

Tufte and Powsner take this to the next level, by summarizing dozens of variables, adding important context (what matters often is not an absolute level but recent trends and long-term status – what is normal for this patient), and leaving the text at the side. A few groups attempted to take this concept and move forward with it, but they aren’t in wide use. It seems like a gem of an idea. Patient records should be easily carried with the team, contain a wealth of information very simply presented, and remove all of the distracting numbers in favor of relative thresholds. Especially interesting is the summary variable for psychosis – quantitative measures for most mental illnesses have been developed but, again, slow to catch on.

Although I have an idea why we haven’t implemented these functions broadly (I am being overly positive about the idea on purpose), I’d be interested in the now and future readers’ comments. Does this concept work? How could we improve upon it? If you think it can’t be successful, why? What other methods appeal to you?

There are no end of articles on Bertillonage, whether because it harkens to eugenics (a trendy area of study), because it occurred during the Victorian era (a trendy era of study), or because the mythology surrounding it marks it as the just-short-of-eureka! moment at which we progressed out of the dark ages of forensic pseudo-science into the enlightened age of scientific methods. At any rate, I am neither interested in Bertillonage for its cred factor, nor its mythological proportions, rather, I am interested in how we have never quite made it away from it after all. The problems posed by Bertillonage are still the problems we struggle with today, especially in informatics.

For those wondering, a quick history is in order. Bertillonage, created by a police clerk named Alphonse Bertillon, was a system intended to identify recidivists (repeat offenders) entering the Parisian prison system. Fingerprinting had not yet been used by police departments, so there was no way to identify a person other than by the name they gave upon being admitted for processing. Bertillon, however, devised a system that would take a very precise series of body measurements – finger length, ear length, head circumference, etc. – and other notations – eye color, hair color – along with a photograph, and file them by a group of the measurements. In this way, a specific card could be looked up very quickly as a subset of the whole. Where before there might be many thousands of cards in the system to sort through, now a person could be measured and looked up based on their measurements…from a group of only a few dozen cards. In other words, Bertillon created an early information retrieval system.

The system worked spectacularly well and quickly made its way across the world – Paris, New York, Boston, London. But this is where problems began to crop up, both in the accuracy and precision of the system. In Paris, the system was scrupulously attended to personally by Bertillon, who painstakingly trained all the measurement takers and had written requirements for exactly how much weight must be placed on a body part as it was measured, how much tension should be used when setting the calipers, what sort of margin of error was permitted. As the system spread, and much to Bertillon’s chagrin, departments changed the methodology and terminology as they saw fit. Where Bertillon might require a precise description of a person, other departments might use terms like “small, medium, and large,” which, of course, were wholly subjective, not mathematical, as Bertillon proscribed. Other departments might drop a requirement for the number of body parts measured, thus lowering the statistical likelihood that a certain set of measurements identified only one person. Most were not careful with the way that measurements were taken, thus ensuring that a person entered into the system by one clerk would have a different set of numbers than if they were later measured by a different clerk.

The system began to unravel, less because of its lack of scientific foundation than because of the difficulty in ensuring consistency in taxonomies, data quality, and accuracy. When fingerprinting broke onto the scene only a few years later, it did not take long for most departments to make the shift from Bertillonage to fingerprinting (though the switch did not happen overnight, it was not particularly protracted, either). Fingerprinting offered the appearance of a system that did not require expertise, something “foolproof” and easy – a system where the statistical differences between people were immutable and didn’t change based on the person doing the measuring. Bertillonage, on the other hand, was cumbersome, required extensive training, and precise methodologies to enjoy the same statistical benefits.

Ironically, many data retrieval systems continue to suffer the same failings today – over 100 years after Bertillonage was widely supplanted by fingerprinting. Developing methods to ensure precision, standards, semantics, data quality, and taxonomical conformance, even when those removed from the development of a system may not understand the importance of such steps, continues to challenge the builders of these systems.