Making Connections
By Jessica Zimmer
Radiology Today
Vol. 26 No. 3 P. 10

How AI-Based Software Changes What Occurs After Imaging

For decades, physicians and other health care providers have struggled with getting patients to adhere to follow-up recommendations after medical imaging. Patients who act on such recommendations usually have better health outcomes. Physicians and other health care providers benefit, too: Their patients fare better, and their financial and litigation risks are lower.

As AI and generative AI (gen AI) have advanced, they have emerged as new tools to encourage and track follow-ups. AI is defined as software that uses models to perform set tasks, such as problem-solving, while gen AI is software that uses generative models to produce new content. AI tools can prevent patients from “falling through the cracks,” even if they change doctors or health care providers. Programs relating to follow-up vary, but many possess the ability to scan for recommendations in radiologists’ text reports, track recommendations referring physicians have made to patients, communicate missed cases, and document patients’ adherence to the recommendations.

“Radiologists could be liable if they make an observation or recommendation in a report that’s ignored by a patient or other doctors,” says Michal Meiri, cofounder and CEO of Agamon Health, a New York-based gen AI medical imaging platform. “The majority of severe diseases that get diagnosed, particularly different types of cancers, start their journey in radiology.”

Meiri adds Agamon Health’s software, Agamon Coordinate, has increased patient adherence to follow-up recommendations. It has also improved communication between referring physicians and radiologists.

“When a doctor knows what the radiologist is recommending, they can align it with their understanding of the patient’s condition. The doctor can then personalize the message about necessary follow-up imaging, treatment, and lifestyle changes. This increases the likelihood that the patient will act on the recommendation,” Meiri says.

How Gen AI Works
Gen AI programs related to recommendations contain large language models, coded programs with algorithms that enable them to understand and generate natural language. As a result, a program such as Agamon Coordinate can review an MRI report, locate the radiologist’s description of concerns—for example, a blood clot—and understand recommendations regarding such issues in an associated report. The program then automatically tracks the follow-up status of each patient and notifies the patient’s referring doctor and/or primary care physician of the need for follow-up.

Gen AI programs related to follow-up recommendations differ from content generation AI tools on the internet, such as Open AI’s ChatGPT. The specialized programs are trained on medical datasets. They are also often able to perform tasks like reminding a doctor to reach out to a patient. Some programs can message a patient digitally via email or text.

Some of these programs also store images in an easy-to-access database. Patients can log in to view their images as they talk to their doctor or scheduling staff. This helps patients view their images and grasp the meaning of the radiologist’s report in real time.

Gen AI programs may involve advanced algorithms and rely on powerful processing chips, but they succeed for a very human reason: The programs get patients more invested in their health. PocketHealth, a Toronto-based company that supports image exchange and patient engagement at health care sites across the United States and Canada, also offers AI-powered tools, including generative AI capabilities, that are designed to support patients in accessing, understanding, and taking appropriate next steps in their care. The platform has a follow-up detection and management solution that identifies when further care may be required. Rishi Nayyar, cofounder and CEO of PocketHealth, says he realized patient buy-in was needed when software programs made it possible for patients to access their medical images and reports.

Connecting With Patients
“Providers found that some patients still did not act on follow-up recommendations, even when they had access to their medical images and reports,” Nayyar says. “Modifying the system to notify patients about these recommendations and track adherence led to an increase in follow-up rates.”

Determining when follow-up recommendations are no longer necessary is also an important consideration. Seetharam Chadalavada, MD, MS, CIIP, FSIR, is CMO for PocketHealth and vice chair of radiology informatics at University of Cincinnati (UC) Health. He supports PocketHealth’s product strategy with clinical insights. UC Health currently uses PocketHealth’s Patient Connect software program to provide patients with access to medical images integrated within their portal. Chadalavada says that gen AI programs help improve visibility and accountability. The software supports follow-up tracking and identifying when follow-up recommendations may no longer be necessary.

PocketHealth also enables providers to gather direct insights from patients about their communication preferences. For example, a patient concerned about recurring cancer may opt to receive text reminders for annual follow-up imaging. PocketHealth’s patient-centered interfaces contrast with software tools that are designed primarily for health systems and providers. A system-focused mindset can lead to gaps in patient engagement and adherence.

Across North America, health care systems are experiencing instability. Political and financial issues have led to hospital closures. This means more shifts in health care access. The disproportionate number of older adults in the population indicates a need to prioritize them. Timely access to screenings and advice could save many lives.

One key area of interest is lung cancer. Early diagnosis of lung cancer saves lives. There are national screening programs that look for it. Yet participation in such programs, or uptake, is quite low: 16%. AI could help by closing care gaps and ensuring patients adhere to follow-up recommendations.

Optellum, a Houston-based software developer, has developed Virtual Nodule Clinic (VNC), a platform for improving the detection and management of lung nodules and cancer. A 2020 study published in BMC Pulmonary Medicine determined that around 65% of lung nodules are discovered incidentally.

Optellum’s Lung Cancer Prediction AI (LCP AI) is a component of the VNC platform. This tool is an FDA-cleared AI/radiomics-based digital biomarker for lung cancer. The LCP AI uses deep learning to produce a patient’s score for lung cancer risk, explains Stephanie Martin, customer success manager for Optellum. Currently, hospitals across the United States use VNC to automatically identify cases for specialist review from radiology reports. From that point on, a practice’s lung nodule team can use Optellum’s dedicated workflow management system to ensure timely follow-up.

Addressing Concerns
Currently, radiologists and doctors are pretty good at identifying and communicating results that might cause acute harm to a patient. “We’re not so good at identifying and effectively communicating things that might harm the patient in a few years,” says Paul Chang, MD, a professor of radiology and vice chair of radiology informatics at the University of Chicago. “AI programs have great potential to help with identifying and documenting incidental findings that are ‘out of band,’ or unanticipated.”

As software companies develop and sell various gen AI programs, often with specific areas of focus, doctors and health care systems are expressing a preference for tools that work with one another. They also want applications that can leverage existing or legacy hospital IT systems, especially with respect to communication.

“Existing ‘homegrown’ hospital incidental finding approaches tend to be onerous, inefficient, and fulltime employee resource intensive. Anything vendors can do to offer solutions to mitigate this problem is welcome,” Chang says.

One change in the landscape that may help encourage patient adherence is familiarity with and access to smartphone applications, which not only applies to younger patients but an increasing number of older health consumers as well.

“Software developers can leverage this more digitally aware population to communicate directly to them,” Chang says. He notes, however, that although direct electronic communication solutions are welcome, there is still a need to augment this approach. “[This is] to address the still significant proportion of patients who do not use electronic-based communication methods or devices,” he adds.

Doctors must also be prepared to take patients’ financial stability and resources into account. “Patients may not have access to transportation, a cell phone, or texting. Their job may not give them the time off. They may just decide the follow-up is not worth it. They may not have the necessary insurance or be able to pay for the care. They may also just forget. Studies have shown that follow-up adherence is worse for economically disadvantaged populations,” Meiri says.

When doctors and health care systems listen to patients, they have a better chance of “meeting them where they are.” This is likely to improve access to care and patient outcomes. For example, if a hospital recognized that patients lacked transportation to medical offices and were experiencing too much pain to take public transportation, it could form a partnership with a local nonprofit. Volunteers could offer quick rides to patients who don’t have vehicles or gas money. If the program was successful, the hospital and the nonprofit could frame the program as a study on patient adherence to apply for a grant.

Next Steps
Faster chips will always increase the speed at which AI programs operate. Other advances will depend on what the software accomplishes. For example, Optellum’s LCP AI relies on computer vision to analyze CT images and produce clinically validated cancer risk scores. Improvements in CT imaging and advances in computer vision may be able to spot new issues, including those that are smaller in size. This might assist with identifying concerns further in advance of cancer development.

Other improvements in software tools may emerge after continuous, long-term exchanges of information between software developers and health care systems. Typically, timing is critical. For example, doctors in a variety of fields want to identify when a noncancerous nodule is at risk of becoming cancerous.

“That is the stage when the cancer is most treatable, and the likelihood of a favorable clinical outcome is highest,” Meiri says. “Screening programs have been shown to save lives and improve clinical outcomes. Breast and lung cancer screening programs are the most well-known types of screening programs. Other screening programs include prostate screening and calcium scoring for heart disease.”

Doctors who utilize AI programs to notify patients of the need for follow-up appointments after screenings could tell software developers how patients in different demographics prefer to receive recommendations. For example, older patients may not be responsive to text messages with follow-up recommendations, possibly because they are not paying attention to smartphone notifications. With authorization from these patients, the hospital could try notifying patients’ family members.

If the secondary method of communication also failed, the hospital could reach out to the software developer to see what adjustments could be made to the software. A potential change may be an automated notice to the specialist with whom the patient has an upcoming in person appointment.

Meiri says health care systems should keep in mind that medical imaging is not the only clinical service that requires follow- up tracking and management.

“Nonimaging follow-up appointments, such as testing, consultations with specialist physicians or primary care providers, and treatment appointments all require follow-up adherence,” Meiri says. “Improving patient adherence to all of these follow-up recommendations can improve patient referrals. They keep patients in network and ensure they return for all forms of follow-up care.”

It is also important to remember that patients are not static entities. Their communication preferences can change. For example, a patient might respond to a text one week but ignore it the following week.

Communicating with patients about their lack of engagement is often a good tactic, especially if the phone call, text, or email contains statements that convey kindness and patience. After communicating the recommendation, it is hard to go wrong with statements such as, “We are here for you. We’re ready to listen and help.”

— Jessica Zimmer is a freelance writer living in northern California. She specializes in covering AI and legal matters.