AI in Medicine: 
The Promise and the Paradox

Barbara A. Crothers, DO ’86, is chief scientist for AlxMed, a cytopathology software company based in Santa Clara, California, and Taipei, Taiwan. AlxMed’s niche is non-gynecologic cytology—all cytology other than Pap tests.

Cytopathology has thus far been an unmet need in digital pathology, Dr. Crothers explains, because unlike surgical pathology, which was “already dealing with two-dimensional images on a glass slide, tissue samples that are a couple of microns thick, and two or three layers of cells,” cytology requires much more scanning. “We compress what we call Z-layers into one image when we’re using a digital scanner, so that the pathologist doesn’t have to use a microscope to focus up and down—they’ll see all of the cell’s features in one flat image.”

AlxMed uses artificial intelligence to deal with the quantitative diagnostic criteria that pathologists have a hard time evaluating “because we’re subjective,” she explains. “For the most part, when we’re doing microscopic work, we’re visually evaluating the size of the nucleus and comparing it to the size of the cytoplasm. But we know that certain cancer cells, for example, have a high nuclear-to-cytoplasmic ratio. And there are some standardized nomenclature systems in cytology for certain body sites, like urine cytology, that specify what that nuclear-to-cytoplasmic ratio should be in order to define that cell as a cancer cell or a potential cancer cell. The computer can do this very well when it’s trained with artificial intelligence.”

Software companies take a whole slide image and overlay it with AI training models that look at particular components on that slide, “in order to do tasks that pathologists do every day,” Dr. Crothers explains. “For example, we have to count the number of mitotic cells in a tumor by examining high-power fields. We know from studies that we’re not very good at it—we don’t get very good concordance, but it’s important. This can be done by AI very well and very quickly.

“More complicated is training the software that this particular cell is probably malignant, which we do by looking at the cell characteristics. In urine, sometimes you’re looking for a needle in a haystack. You may not have many abnormal urothelial carcinoma cells in a urine specimen, but AI can pick them up a lot quicker than any of us can looking at the slide, and it can also categorize them quickly. The software we’ve developed presents those cells to you on a screen as thumbnail images, with the statistics for each cell, so you know exactly what that cell’s characteristics are.”

The same model can be transformed slightly to fit other specimen types, such as thyroid specimens. Down the line, says Dr. Crothers, “we’re going to be looking at lung and probably other systems, to create an entire suite of AI software modules.”

Much of AlxMed’s current research is in preparation for an FDA submission, since medical devices can be used in clinical practice only with FDA approval. “That hurdle is very high,” says Dr. Crothers, “and there’s only been one approved AI in cytology to date, just this past year. The FDA is very helpful in giving you guidance on what kind of data they’re looking for. An FDA validation study is usually a very large study that shows proof of concept and the device’s reliability in a medical setting.

“The device doesn’t make the diagnosis, the pathologist does. We have extensive quality assurance programs in pathology, and we’re checking each other all the time. But we could rely on AI to help check us as well, instead of needing another set of human eyes—maybe have AI do a lot of that back-end quality assurance work that we do every day.”

In that case, will pathologists lose their jobs?

“I saw this same concern arise with the advent of automated Pap tests,” says Dr. Crothers. “Screening guidelines changed; test volumes dropped, and automation reduced the time required for screening Pap tests. We thought cytotechnologists, who screened most of the Pap tests, would lose their jobs. But meanwhile, there was an unmet need in the lab for small biopsy, fine-needle aspiration specimen quality evaluation because of the implementation of robotics and new methods of collecting specimens and cells. We needed cytotechnologists to go to these procedures to evaluate the adequacy of these specimens. Very few cytotechnologists lost their jobs.”

What’s more, she points out, “there continues to be a decline in the number of individuals who are doing cytotechnology. Schools are not putting out enough technologists to meet the need, and the shortage is worldwide. The same is true for pathology, and honestly, the same is true for medicine. We’re at a crisis point; we don’t have enough individuals to do the work. AI could free us up from some of the time-consuming manual tasks.”

Besides her involvement in AlxMed’s research studies, Dr. Crothers is responsible for developing the training materials and for training investigators and future customers on the use of the software. “If I’m involved in the process,” she says, “I can help to fine-tune how it will be helpful to pathologists in their day-to-day practices.”

As chief medical information officer at Main Line Health, John Potts, DO, RES ’00, FAAFP, is responsible for facilitating the electronic medical record (EMR) strategy, implementation, optimization and data analytics for users throughout the Main Line Health Network. This means ensuring that people have the technology they need in their roles and the knowledge to use it—and AI increasingly is central to that technology.

About the concern that AI is going to take people’s jobs, Dr. Potts says, “AI offloads tasks, whether from a physician, nurse, medical assistant or medical biller. I suppose if you offload enough tasks for someone’s specific role, you might be able to repurpose them to a different role. But I don’t see artificial intelligence replacing people. In fact, as baby boomers get older, and sicker, and Gen X is starting to head into the retirement years, we don’t have enough of Gen Z coming out of schools to replace folks that are retiring. We can’t hire our way out of this.”

An example of AI providing diagnostic assistance is Main Line Health’s use of a platform that reviews CAT scans to not only look for large vessel strokes, but to “immediately alert our stroke team by phone to go see that patient, and it tells them where that patient is,” says Dr. Potts. “It also alerts our radiologists and prioritizes that film to be read immediately. If the patient truly is having a stroke, then we can begin treatment right away. And that’s allowed us to decrease our times to treatment for the appropriate patients.” Besides neurology, the platform is being used in cardiology and pulmonology.

Clinician assistance can take the form of AI generating replies to questions on patient portals. It turns out that “AI is far more verbose than our clinicians,” says Dr. Potts. “Studies have shown that patients often like the AI replies—‘Look how much time the doctor spent writing back to me’—but a long reply might encourage a patient to keep asking more questions over the portal as opposed to coming in for a visit. That’s always been a feature of having patient portals, of course, but we had to work with our AI and prompt training to dial down the length of the replies.”

What’s more, the clinicians might say that a reply doesn’t capture their “voice.” Dr. Potts recalls, “We once had a patient call the office to ask, ‘Is the doctor feeling okay? Their reply to me doesn’t sound like them.’ The AI we use is a generic model—a HIPPA-compliant version of GPT-4. Training AI so it can mimic clinicians is extremely expensive and very challenging. But companies are pursuing this, and our development team is working on it as well, so that the AI would train off a model that would be tailored to individual clinicians.”

Whether you’re a mid-size health system like Main Line Health or a larger network, says Dr. Potts, the key question is “how do you scale whatever you’re bringing to the clinical side. For example, we’re looking to create an innovation room—a patient room that brings in all the new technologies to see how they work with the other equipment, what the workflows are like. We want to bring in doctors and nurses to test innovations and determine if they’re ready for rollout on a broad scale.”

Dr. Potts predicts that the hospital is going to radically change in the next three years for patients and for staff and become a “smart hospital room.” “Technologies using cameras can take a patient’s vital signs, contactless—pulse rate, heart rate, temperature, and we’re told that blood pressure is coming in the near future. A nurse or care tech won’t have to interrupt the patient, whether they’re with family during the day or sleeping at night. No one will have to manually enter those vital sign numbers into the EMR.

“AI in that same camera can detect whether the patient is a fall risk, and if so, alert somebody on the floor. It can show us if a patient is at risk for developing pressure ulcers, or is smoking in the room and at risk of causing a fire, or if somebody has brought a weapon into the room. There’s lots of use cases for AI in the room, not only for the patient, but for the hospital staff as well.”

Dr. Potts points out that every specialty can benefit from AI. “Think dermatology, for example. AI models can do as well as a person looking at a lesion to determine if it’s malignant or not. Precision medicine and human genomics—they’re sequencing patients’ DNA and then loading it back into the EMR. AI can look at a humongous amount of data and help predict for us what cancer treatments are the best options for a particular patient based on their genomics. That is really exciting, and that day is coming.”

As a pharmacy solutions analyst at CaryHealth, Prerak Adhuria, PharmD ’17, explains, “I’ve been working very closely with marketing sales as well as development enhancements of products within our company that use AI—improving our tech and seeing the patient outcomes in the digital pharmacy experience—as well on the patient side and on the health plan side, so payers and health systems can make better decisions. My role is essentially to be a bridge that supports the C-suite, connecting my pharmacy background with the technology background I’m learning and then tying it all together.”

Dr. Adhuria is currently the project manager for all CaryHealth conferences—for the pharmaceuticals industry, for payers, for providers—related to Clair: Clinical AI Reference, a tool focused on fast, precise information retrieval. Besides attending some of these conferences, he vets them and determines how the team will reach out to attendees. Clair, which is designed for quick lookups of medical guidelines, evidence-based practices and other clinical data without extensive patient-specific integration, gives users answers within seconds, says Dr. Adhuria, rather than the minutes needed by other resources.

Clair’s answers are geared toward whoever’s asking it a question. The app allows retail pharmacists, physicians and other professionals who see patients throughout the day to answer patient questions right away, face-to-face, rather than making an extra trip to look something up. “A professional can ask Clair about a medication’s effects on blood pressure or drug interactions,” explains Dr. Adhuria. “But a lay person could ask Clair if it’s OK to be taking three medications together. Meanwhile, the admin team could ask about ICD-10 codes for billing purposes.”

Some people think of AI as “an engine out there that does its own thing,” Dr. Adhuria says, “but with proper guardrails and monitoring, it makes people’s job’s easier, allowing them to focus on the patient’s needs.” Dr. Adhuria especially loves to connect with students, and at a recent conference of the Academy of Managed Care Pharmacy, “I got to see students’ reactions,” he says, “and it was blowing their minds.”

OneDash, another CaryHealth offering, is a population health tool: an AI-driven clinical automation platform designed to identify and automate the closure of care gaps. A plan can, for example, immediately see if any patients with diabetes are not on a statin, or pull all patients in a particular age range that have been taking a particular medication over the past three months. Previously, that might take the health plan a couple of weeks. Now, says Dr. Adhuria, “the turnaround is 10 or 20 minutes. And that allows decisions to happen faster.” Automations save time, too, not only by sending faxes to doctors’ offices but by making AI-assisted calls to patients “that sound like a normal conversation,” says Dr. Adhuria. He notes that a health plan in the DC/Maryland/Virginia area has recently been able to see financial savings, improved patient outcomes and increased patient adherence rates.

“A doctor can do their job,” he continues, “without having to handle prior authorizations submitted by pharmacies, without having to reach out to the insurance company. It was a mess, and now all this is automated. Doctors love this because now they can spend more time taking care of patients; pharmacists love this because now they can talk to patients who come to the pharmacy to get their prescription.”

Before he graduated from PCOM, Dr. Adhuria worked for almost a decade as a technician and intern with CVS. After he graduated, he did independent pharmacy work in a small town, with a very small patient base—home care, assisted living services, packing medications in a bubble pack, delivering those medications at 8:00 p.m. if I had to. I knew people had no other means of getting what they needed, and I thought, ‘There has to be a better way.’ Next, working with a mail order retail pharmacy, he learned “how to improve on efficiencies and get 100 percent of our medications out every day without sacrificing quality. So I went from delivering medication myself to someone’s house to being able to ensure they could have it mailed to them before their medications run out.”

Now, with CaryHealth, Dr. Adhuria works with a digital pharmacy that is licensed in 50 states and “offers phenomenal patient experience. Whether you need to refill a medication, or discontinue a medication, or want to check on something, you can just do it from your mobile app.”