Tackling the Problem of Prescription Drug Diversion

Contributed Commentary by David Brzozowski Jr.

May 14, 2021 | Although 9 in 10 patient care professionals handle prescription drugs responsibly, the 1 in 10 who steal and abuse them are putting themselves at mortal risk and forcing patients to endure untreated pain. Healthcare organizations comprised of hospitals, pharmacies, clinics, emergency centers and other settings are doing what they can to detect and prevent prescription drug theft or drug diversion. The effort, however, typically entails a lot of manual analysis, and for the most part, the results have been disappointing.

Effective drug diversion prevention needs to consider all fronts throughout the medication lifecycle, from wholesaler procurement and delivery, central pharmacy inventory management, pharmacy distribution to point of care, orders and their fulfillment to nurses, and, ultimately, administration to the patient. Obtaining insight across this continuum and making sense of the data between these disparate systems is daunting and, in most occurrences, out of grasp for the overburdened health organization. For this reason, effective drug diversion prevention should be based on power artificial intelligence. Done right, drug diversion prevention powered by AI can save lives, improve patient care, reduce health care costs, and minimize wasted effort.

Here are the challenges health organizations are tackling to get there:

Data evolution: The conundrum for hospital systems that are attempting to enlist AI in drug diversion prevention involves training the AI system. With the evolution of informatics in healthcare, organizations are constantly updating and replacing their information systems to stay relevant with digital transformation. As a result, the historical data lineage left behind from these systems becomes fragmented and, in most cases, unusable.

The possession of such data is merely a foundation for future potential. To leverage this data for effective insight, there must be mechanisms built to normalize and enrich the data, accounting for every modification that’s occurred throughout its historical existence. Such processes if conducted manually are exhaustive, and thus, the use of AI in data purification is becoming more prevalent.

Data depth: When evaluating a snapshot of an organization’s current data you may draw some conclusions, however these conclusions come with uncertainty. Without significant depth in the sample set being used, cultural and procedural changes can skew analysis leading to possible false accusations. As an example, with the introduction of COVID-19 in early 2020, there has been a major decline in the amount of elective surgical procedures. As a result, data captured during this time will reflect this cultural change. Thus, when the impacts of COVID-19 pass, the data will again be impacted. Each dramatic change in the data skews the sample resulting in inaccuracies and ultimately a lack of confidence in the conducted analysis. For this reason, leveraging a sample with a broad historical scope spanning years instead of months ensures confidence in the resulting analysis as cultural and procedural shifts impacting the data are accounted for. Without handling data in large quantities that have accumulated over time, the AI cannot become truly intelligent and, therefore, accurate.

Fortunately for today’s emerging AI efforts (though not, perhaps, for society), opioid diversion was a concern long before widespread abuse hit the headlines. At Medacist, we have data from 21 years ago; and there are increasingly detailed drug-flow records for the intervening two decades across thousands of hospitals. If aggregated and continuously updated, such a collection can provide AI developers with a solid baseline for what normal prescription drug patterns look like.

Normalizing data: To amass such data from disparate healthcare systems and their pharmacies, technologists need to gather apples with apples. The problem is that one vendor’s pharmacy information system, automated dispensing system (ADS), or electronic health record system typically uses different data formatting from another’s, creating a Tower of Babel that someone needs to dismantle. For example, one vendor’s system may store a patient name in four fields—first name, middle name, last name and suffix—while another system may use a single field to store the name.

Similarly, technologists need to reconcile brand-name drugs and their generic equivalents (there’s no reason to analyze them separately) as well as different units of measure for the same drug e.g., milligrams/micro kilograms. Then there is standardizing employee data on time & attendance, drug access and waste. Technologists are constantly working to resolve incongruities like these to create “canonical” data models—and ultimately one version of the truth.

Human input: AI is a broad term that signifies a computer system assisting a human in decision making. It’s always impressive, but in the best case, the software is constantly improving its performance—the machine is learning. Cutting-edge drug diversion solutions do allow the machines to learn with just a modicum of assistance from humans. Every time the system flags a clinician for anomalous behavior, a trusted user tells the software whether it was right or wrong. This user intervention is as easy as (and similar to) confirming Facebook’s accuracy when it guesses at the identity of a friend in a photo you’re posting. Drug diversion systems ask users to click yes or no to indicate whether anomaly is a confirmed diversion, a false positive, or in rare events, a false negative. It’s embedded in the workflow.

Automatic updating: The most advanced drug diversion systems bring real-world, as-it-happens data streams on drug flow back into the system as often as possible: up to six times a day if there’s enough data to justify the activity. This way, the historical baseline is consistently refined. User feedback, however, is integrated in real time. And if a behavioral change was collected through user feedback, the system would account for such feedback in real-time triggering a re-evaluation of the activity in question yielding updated analysis.

These are some of the most important challenges that, as they’re overcome, will help the nation’s healthcare systems ensure that prescription drugs go directly to those who need it most. Diversion detection is not a practice that works well without automation and without constant learning by machine, together with some basic feedback from human users.

The result of this machine learning approach will be healthier patients and care professionals, and health-system efficiencies that can improve access to high-quality care for all.


David Brzozowski Jr. is Chief Technology Officer of Medacist Solutions Group, a leading innovator in drug diversion analytics. He can be reached at djrbrzozowski@medacist.com.