How will AI change healthcare?

Doctors are understandably skeptical of claims that artificial intelligence will transform medicine. Remember the misleading claims that radiologists may soon be obsolete, the years of annoying automatic pop-ups of electronic medical records, and the almost useless introduction of IBM’s Watson. But the development of new large-scale language models may actually live up to the hype. GPT-4, the latest, largest and most functional model developed by OpenAI, has passed many of his AP exams and various professional certification exams (including Sommelier exams) even without training. pass the

In medicine, the GPT-4 effortlessly passes the series of exams that aspiring doctors must pass to become licensed, with an 83% accuracy rate. (The minimum passing score equates to approximately 60%.) We also achieve excellent results with board resources designed to prepare physicians for the American College of Physicians exams. Compared to physician responses to 195 patient questions, GPT-4 responses were rated higher for empathy by a blinded team of medical professionals, although they were not rated for accuracy.

While no state has yet licensed GPT-4 to independently practice medicine, the tool appears poised to change medical practice in many ways. However, given this particular model’s propensity to “hallucinate” misinformation, and given its outdated knowledge and manufacturer-issued warnings, a human physician probably needs to supervise its use. prize. This is called “Doctor In”. -The Loop. ” Still, GPT-4 is poised to streamline healthcare. Providing enough context, explicit and detailed instructions (or “prompts”) to the tool, and finally some internet access, can improve performance compared to an already excellent baseline. I have.

Let’s consider some hidden uses first. GPT-4 has the potential to speed up administrative workflows by automating the completion of pre-approval forms for drugs and the filing of appeal letters to insurance companies that deny treatment.When it comes to healthcare systems, more insight into opaque medical records will be possible: says health tech entrepreneur Will Manidis put itAI models make “data computable” by transforming messy patient records into a more usable format.

For physicians overwhelmed with messages from patients, autoresponder drafts and helpful summaries would be a welcome addition. Pilot projects are already underway at many medical institutions. Meanwhile, a virtual “digital scribe” combines voice transcription technology with her GPT-4 to listen to patient and doctor visits and automate note-taking. The banknotes are then automatically checked for possible billing codes, resulting in optimal returns. While this is a boon for doctors, it is unlikely to be welcomed by payers such as insurance companies and Medicare.

Patients who struggle to understand complex terminology or who lack English proficiency benefit from the ability to translate medical records into understandable formats in different languages. Eventually, patients and doctors may be able to ask medical records questions and receive contextualized answers, much like financial technology company Stripe does in its developer documentation.

Some uncertainties remain. What impact will reducing some of the healthcare transaction costs have on overall system costs? To predict , we need a theory about which transaction costs will fall and which others will rise.” One mechanism for controlling health care costs is “utilization management.” For Medicare Advantage insurance plans in particular, this means protecting the supply of expensive drugs, surgeries, and tests by hiding them behind complicated insurance company paperwork. Viewed this way, high transaction costs can be positive from a system perspective. Patients who need really expensive treatments can eventually get them, but cheaper alternatives are tried first, thus lowering overall costs. In practice, however, this can be a very frustrating ordeal. Horrifying stories occasionally emerge of desperately needed treatment being denied and delayed.

What if you could create a perfectly formatted pre-approval and the cost of subsequently disputing an insurance denial would be close to zero? The first order effect could be cost savings. Staff time spent on insurance-mandated paperwork is significantly reduced, freeing up time for other tasks and improving the patient experience. However, inferring second-order effects requires some context. Approximately 94% of pre-submitted approvals are approved. Only 11 percent of rejected applications are subsequently appealed, of which approximately 80 percent lead to partial or full grants. The net effect will likely be a slight increase in the percentage of pre-approvals approved on the first go-around, followed by a much higher number of appeals and approvals. Insurers could also use systems like GPT-4 to audit pre-approvals as an alternative to the crude automated systems currently in use by some.

AI is only getting better. GPT-4 performs well even in the presence of task-specific training and real-time access to existing medical databases and various professional guidelines. As new business models and regulations are developed with GPT-4 and its successors in mind, even greater changes are likely to occur in healthcare.

Photo credit: Jakub Porzycki/NurPhoto, Getty Images