Using AI to ID Osteoporosis: A Medico-Legal Minefield?

Could an artificial intelligence (AI)–driven tool that mines medical records for suspected cases of osteoporosis be so successful that it becomes a potential liability? Yes, according to Christopher White, PhD, executive director of Maridulu Budyari Gumal, the Sydney Partnership for Health, Education, Research, and Enterprise, a research translation center in Liverpool, Australia.

In a thought-provoking presentation at the Endocrine Society’s AI in Healthcare Virtual Summit, White described the results after his fracture liaison team at Prince of Wales Hospital in Randwick, Australia, tried to plug the “osteoporosis treatment gap” by mining medical records to identify patients with the disorder.

‘Be Careful What You Wish For’

White and colleagues developed a robust standalone database over 20 years that informed fracture risk among patients with osteoporosis in Sydney. The database included all relevant clinical information, as well as bone density measurements, on about 30,000 patients and could be interrogated for randomized controlled trial recruitment.

However, a “crisis” occurred around 2011, when the team received a recruitment request for the first head-to-head comparison of alendronate with romosozumab. “We had numerous postmenopausal women in the age range with the required bone density, but we hadn’t captured the severity of their vertebral fracture or how many they actually had,” White told the Medscape Medical News. For recruitment into the study, participants must have had at least two moderate or severe vertebral fractures or a proximal vertebral fracture that was sustained between 3 and 24 months before recruitment.

White turned to his hospital’s mainframe, which had coding data and time intervals for patients who were admitted with vertebral or hip fractures. He calculated how many patients who met the study criteria had been discharged and how many of those he thought he’d be able to capture through the mainframe. He was confident he would have enough, but he was wrong. He underrecruited and could not participate in the trial.

Determined not to wind up in a similar situation in the future, he investigated and found that other centers were struggling with similar problems. This led to a collaboration with four investigators who were using AI and Advanced Encryption Standard (AES) coding to identify patients at risk for osteoporotic fractures. White, meanwhile, had developed a natural language processing tool called XRAIT that also identified patients at fracture risk. A study comparing the two electronic search programs, which screen medical records for fractures, found that both reliably identified patients who had had a fracture. White and his colleagues concluded that hybrid tools combining XRAIT and AES would likely improve the identification of patients with osteoporosis who would require follow-up or might participate in future trials.

Those patients were not being identified sooner for multiple reasons, White explained. Sometimes, the radiologist would report osteoporosis, but it wouldn’t get coded. Or, in the emergency department, a patient with a fracture would be treated and then sent home, and the possibility of osteoporosis wasn’t reported.

“As we went deeper and deeper with our tools into the medical record, we found more and more patients who hadn’t been coded or reported but who actually had osteoporosis,” White said. “It was incredibly prevalent.”

But the number of patients identified was more than the hospital could comfortably handle.

Ironically, he added, “To my relief and probably not to the benefit of the patients, there was a system upgrade of the radiology reporting system, which was incompatible with the natural language processing technology that I had installed. The AI was turned off at that point, but I had a look over the edge and into the mine pit.”

“The lesson learned,” White told Medscape Medical News, is “If you mine the medical record for unidentified patients before you know what to do with the output, you create a medico-legal minefield. You need to be careful what you wish for with technology, because it may actually come true.”

Grappling With the Treatment Gap

An (over)abundance of patients is likely contributing to the “osteoporosis treatment gap” that Australia’s fracture liaison services, which handle many of these patients, are grappling with. One recent meta-analysis showed that not all eligible patients are treated and that not all patients who are treated actually start treatment. Another study showed that only a minority of patients — anywhere between 20% and 40% — who start are still persisting at about 3 years, White said.

Various types of fracture liaison services exist, he noted. The model that has been shown to best promote adherence is the one requiring clinicians to “identify, educate [usually, the primary care physician], evaluate, start treatment, continue treatment, and follow-up at 12 months for to confirm that there is adherence.”

What’s happening now, he said, is that the technology is identifying a high number of vertebral crush fractures, and there’s no education or evaluation. “The radiologist just refers the patient to a primary care physician and hopes for the best. AI isn’t contributing to solving the treatment gap problem; it’s amplifying it. It’s ahead of the ability of organizations to accommodate the findings.”

Solutions, he said, would require support at the top of health systems and organizations, and funding to proceed; data surveys concentrating on vertical integration of the medical record to follow patients wherever they are — eg, hospital, primary care — in their health journeys; a workflow with synchronous diagnosis and treatment planning, delivery, monitoring, and payment; and clinical and community champions advocating and “leading the charge in health tech.”

Furthermore, he advised, organizations need to be “very, very careful with safety and security — that is, managing the digital risks.”

“Oscar Wilde said there are two tragedies in life: One is not getting what one wants, and the other is getting it,” White concluded. “In my career, we’ve moved on from not knowing how to treat osteoporosis to knowing how to treat it. And that is both an asset and a liability.”

Marilynn Larkin, MA, is an award-winning medical writer and editor whose work has appeared in numerous publications, including Medscape Medical News and its sister publication MDedge, The Lancet (where she was a contributing editor), and Reuters Health.