As you and your pharmacy staff wade through hundreds of e-prescriptions, doing manual data entry for most of them while more pile up behind them, do you ever think, “Why can’t ChatGPT do something that actually matters to me?”
Certainly, generative AI has some amazing capabilities — as well as some confounding implications.
I’ve taken the current iterations of ChatGPT and Google Bard for a spin, and I am both impressed and deeply skeptical. I’m impressed by generative AI’s ability to create accurate summaries of articles nearly instantaneously or suggest new, engaging headlines (like for this article, actually). I’m skeptical of how the hype overpromises what these tools can achieve because overpromising and underdelivering could harm pharmacies and their patients.
After all, who hasn’t lost confidence in AI’s ability for mission-critical work every time we use talk-to-text on our cellphones and see so many goofy mistakes?
Admittedly, my skepticism about AI is ironic, given that developing machine learning for the healthcare industry has been my primary career focus for over 12 years. Before AI was an everyday buzzword, some of us working on this technology in healthcare realized there had to be a distinction between AI intended for general use and clinical-grade AI, which has the precision and accuracy necessary for consistent, clinically validated results.
WHAT IS CLINICAL-GRADE AI?
For clinical-grade AI, it’s about the proof, not the hype. When evaluating options for your pharmacy, ask for the metrics.
It may help to think of clinical grade this way: A surgical drill is a far more precise tool than what you can buy at any home improvement store. The surgical drill is considered clinical grade, and the other is for general use. Both make holes and are called a “drill.” In many ways, AI can be thought about in the same way. Only a few companies today make AI tools that are clinically vetted for use in healthcare. Most others develop AI tools that range from theoretically cool but untrustworthy to little more than nifty entertainment. Any confusion between the two types of capabilities is generally due to overhyped marketing and wild expectations.
Here’s why it matters: When a tool is clinical grade, you can have confidence in how it will perform in a healthcare setting where lives are at stake. For example, if I needed surgery, I should be confident that my surgeon would use a clinical-grade surgical drill rather than a drill bought at a home improvement store for a tenth of the cost.
PLANNING FOR SUCCESS WHEN AI FALLS SHORT
For the safety of your patients, you must know what happens when the AI doesn’t perform according to plan. When evaluating AI for your pharmacy, ask what happens when the needs go beyond the AI’s abilities.
Do you remember back in 2013, when Elon Musk announced that we would have 90% self-driving cars by 2016? It was exciting to imagine, but I was doubtful. While it made theoretical sense that a computer could eventually handle most routine driving situations, I couldn’t get one thought out of my mind. In an instant, driving a car can go from routine and orderly to unpredictable and deadly.
And unpredictable is not the realm where AI thrives. For safety, there must be a way for the computer to hand off responsibility to the human driver with more than enough time to allow the human to evaluate the situation, then react quickly and appropriately. After all, lives are at stake.
The same is true for using AI in healthcare as in self-driving cars. What happens when the AI can’t do the job? Does it create an unsafe situation, like a self-driving vehicle handing off control to a human driver with just milliseconds to react? A hallmark of clinical-grade AI is that it performs as “augmented intelligence” in partnership with clinicians. The system is engineered for humans to make all final decisions, but the process is faster and safer via the leg work the computer did to assemble the facts.
It’s like how pilots and co-pilots work together. When you fly commercially, you will notice that there are generally two pilots in the cockpit. This isn’t because flying an airplane is so complicated that it takes two people. It’s a safety measure because a failure at 35,000 feet could have catastrophic results. Together, the pilot and co-pilot serve as a “safety multiplier.”
While one is flying the airplane, the other is watching for potential problems and can jump in when needed. It would take both the pilot and the co-pilot to make the same error simultaneously for a tragic event to occur. That rarely happens, and as a result of this approach, together with many technical innovations, aviation has a superior safety record in comparison to any other mode of transportation, according to the International Air Transport Association.
Clinical-grade augmented intelligence can serve the role of co-pilot in the pharmacy. It can recognize when it cannot safely process a deeply complex transaction. In these occasional instances, the AI will require a manual review by pharmacy staff.
An escalator can illustrate this point. If the motor stops, the escalator becomes a stationary set of stairs everyone can use. When clinical-grade AI encounters a situation it cannot solve, it safely stops that action and flags it for a safe handoff to clinicians to decide.
The clinical-grade AI does the tedious, time-consuming work a person typically does, with the “success from failure” plan for clinician verification. This pilot/co-pilot approach to AI, as augmented intelligence, has proven to be a force multiplier while simultaneously being a safety multiplier as well.
THE PROOF IS IN THE METRICS
Clinical-grade AI has been used as part of the medication reconciliation process by hospitals in the United States since 2015 and in Canada since 2018. Today, hospital clinicians using top electronic health record (EHR) systems are using clinical-grade AI for more than 24 million medication transactions daily. As a result, 93% of medications are reconciled more quickly, with fewer keystrokes and data entry errors. Here are some proof points:
- A 400-bed community hospital and Level II trauma center in southern Massachusetts performed 14% more medication reconciliations and reported 23% fewer medication safety events.
- A major health system in North Carolina avoided 20 million keystrokes, and staff satisfaction increased by nearly 40% within the first 90 days of implementation.
- An eight-hospital system in Pennsylvania converted over 5.5 million medication records from legacy EHR systems, saving five to seven minutes per patient record, with nearly a 90% success rate.
Today, armed with the right questions to separate the proven benefits from the hype, pharmacies can choose an AI partner to increase productivity, avoid keyboard errors, and free pharmacy staff for more meaningful activities, including patient care. CT
David Sellars, C.Eng., serves as principal, product innovations, at DrFirst. He has over 20 years of healthcare experience, with a deep emphasis in big data, AI, healthcare interoperability, programming, and systems optimization. The author can be reached at dsellars@drfirst.com.