Using data to build proactive drug error prevention models

By | August 25, 2015

Data is variously described as the oxygen of the digital economy or the new raw material of the 21st century.“-Nigel Shadbolt

There are more than a few issues with today’s medication order entry systems. However, in this post I want to focus on only two.

First, alert fatigue. As a pharmacist that has entered his fair share of orders I can tell you that alert fatigue is real. Order entry systems, including CPOE, are designed to indiscriminately alert users of every possible problem associated with the patient’s profile and the order being entered. When entering orders for a patient with complex medical conditions, this can become a bit frustrating because a majority of these alerts are of little to no value. After a while you begin to blow through alerts because so many are simply a waste of your time. Unfortunately, when this happens you will occasionally miss something important. It happens.

Second, the “perfect medication error”.(1) This occurs when a physician inadvertently utilizes CPOE to order the wrong medication for a patient – or the right drug for the wrong patient – but the order meets all the necessary checks and balances to end up on the medication profile, i.e. no allergies, meets all appropriate dosing parameters, there are no drug-drug interactions, labs are fine, and so on. This is an issue that appeared on my radar while performing an FMEA for a CPOE implementation when I was still working as an IT pharmacist.

The issue with excessive, inappropriate alerts during order entry is real, and it’s been a thorn in the side of healthcare providers for quite some time. Fortunately the growth of data analytics and business intelligence has given rise to more potential solutions than ever before. Work such as that being done by Dr. Criag Rusin at Baylor College of Medicine is a good example. Dr. Rusin presented on real-time clinical decision support at the ASHP Summer Meeting in June. The work focuses on using machine learning to provide real-time, patient-specific feedback to the end user, minus all the unnecessary information. Promising work, but how long it will take to become mainstream is anybody’s guess.

Another promising development in this area comes by way of Dr. Gidi Stein, Professor of Medicine and Molecular Engineering at Tel Aviv Universityis. Dr. Stein’s work focuses on utilizing a big-data software platform called MedAware that integrates with a hospital’s EHR to detect prescription errors before they happen. According to an article in Modern Healthcare: “[MedAware] draws from patterns in millions of patient records to flag medication-order outliers. If a physician chooses a drug that doesn’t match any condition in the patient’s record or diverges from how other patients with similar histories have been treated, the discrepancy is flagged. The system blocks the drug order until the doctor confirms its accuracy or cancels and re-enters the order. Typically, EHR alert systems are programmed to spot dangerous drug interactions, higher-than-normal doses and duplicate prescriptions. But these systems have not used large volumes of aggregated data to determine in real time the likelihood that the wrong drug was selected for a particular patient.”

Both real-time clinical decision support utilizing machine learning that can adjust on the fly, and predictive analytics that can preemptively stop order entry errors are promising. I would like to see both implemented sooner rather than later. One has to wonder at what point order verification performed by a pharmacist will become obsolete? Given the potential of the work being done by Dr. Rusin and Dr. Stein, it’s easy to believe that pharmacists may finally be able to move out of that role into something with more substance.

(1) This actually happened during one of my mothers hospital admissions. I caught the error only after reviewing her medications on discharge.

One thought on “Using data to build proactive drug error prevention models

  1. Ray Vrabel, PharmD

    Unintended Consequences of CPOE:

    The use of real-time big data, and heuristics will certainly be how to ensure the accuracy of medication therapy decisions in the future. The pharmacist’s role at that time will be to “handle the outliers” (i.e., those orders that cannot be validated by the therapy checking software). That will be far more interesting (and satisfying) to the pharmacist, given that with standardized order sets, most orders are correct to begin with. Not all, but most.

    Both of the problems you identified, order fatigue and wrong drug selection by the prescriber, are both unintended consequences of poorly designed physician order entry systems (CPOE).

    Most CPOE systems are very similar in structure to how pharmacists and pharmacy technicians have been entering medication orders into computers since that 1970’s. It’s no surprise that these “product-based” systems don’t work for physicians who do NOT think in terms of a “pharmacy product” (i.e., how a drug comes from the manufacturer; package size, dosages available, etc.). Physicians think in terms of the “therapy” they want for their patients, not the products they want to have the pharmacy send up to the nursing unit.

    My analogy for this is like dining in a restaurant. If a physician used the “Therapy model” to order their dinner, they would merely say, “I would like a sirloin steak, medium-well, with a baked potato, no butter, with asparagus. In other words, they just ordered what they “wanted for dinner”.

    Using the “CPOE model” in the restaurant, they physician would be forced to indicate the precise number of ounces the steak was cut to, was it farm raised, how long was the steak aged, etc. For the potatoes, the physician would have to specify the exact size of the potato, the color of the skin, and whether it was once baked or twice, etc. The asparagus component would be equally arduous (e.g. size, length, color, chewy vs firm, etc).

    It’s no wonder physicians don’t like most current CPOE systems. The optimum CPOE should allow the physician to think in terms of therapy, not products. The product piece should be handled by the CPOE, but it should be behind the scenes, with the pharmacy staff managing the translation of the therapy order into products. The Pharmacy staff are in charge of the “kitchen”, interpreting the therapy orders received (i.e., the dinner request) and preparing and dispensing the products (i.e., the meal) for administration to patients.

    I remember the “good old days” of medication order review, when you interpreted what the physician wrote on the physician order sheet of the patient’s medical record. When a physician was ordering, they were ordering “everything” for the patient, not just drugs. That meant that a given order sheet might include medications, lab tests, radiology tests, microbiology requests, consultation requests, dietary, etc. Now, sometimes it took a little searching to find the medication order on the sheet of paper, but what it did that most CPOE systems do not do well is present the medication order IN THE CONTEXT of what else is going on with the patient. The pharmacist could “see” the thought pattern (good or bad) of the physician regarding the desired therapy for the patient and using their clinical judgement the pharmacist could better assess the appropriateness of the ordered medications. The chances of a physician writing a totally different medication than what they were thinking is virtually impossible (your second unintended CPOE consequence above), and if they did it would stand out like a sore thumb to the pharmacist reviewing the complete order sheet.

    Fixing this problem should be accomplished in steps:

    (1) Develop a CPOE system from scratch that is designed to meet the needs of the physician, not the pharmacist (i.e., therapy-centric, not product centric).

    (2) Develop the backend of the CPOE system to convert the therapy-centric ordering of physicians into product actions for the pharmacy.

    (3) Present the CPOE orders that the pharmacist sees for verification along with other pertinent, relevant patient clinical information (e.g., labs, microbiology, etc.) allowing the pharmacist to see the drug requested in the context of what is going on with the patient.

    (4) Develop the big data/heuristic systems to streamline the verification of medication therapy orders so that the pharmacist only has to handle the outliers.

    By the way, I would like to see all of this within the next year…!!!

Leave a Reply

Your email address will not be published. Required fields are marked *