NEW YORK – Artificial intelligence may rule the world one day, but can it improve disease diagnosis?
If the pundits are to be believed, all forms of AI technology will pervade every aspect of our lives soon, but when it comes to healthcare diagnostics, its application has been relatively limited so far, even if some predict that the technology could redefine the whole healthcare experience. To date, AI in healthcare has been used for a variety of purposes, including scanning electronic health records for data, as well as to analyze data in clinical trials to find appropriate patients for experimental therapies. In telemedicine, AI technology can allow doctors to make real-time decisions driven by data.
For now, the use of AI in diagnostics has largely been in the pathology space, and according to Emily Volk, president of the College of American Pathologists, AI "won't necessarily replace pathologists, but pathologists who use AI and use it well will replace pathologists who don't."
In spite of some skepticism from CAP members about the technology — mainly around the risks of off-target results and concerns that pathologists will be replaced by the technology — the organization has been pushing pathologists to at least be comfortable with pattern recognition software that has gotten a toehold in the field.
Volk likened the current situation to climbing Mt. Everest. Investments into equipment will be required as will validation of tools that will eventually pay off in improved efficiency and patient outcomes. For now, AI in pathology remains at the base camp of its Everest ascent, she said.
Regardless, diagnostics developers are fervently exploring opportunities to incorporate AI and machine learning into new tools so that they can identify new targets and tell when something has gone wrong in the pre-testing process.
Karl Hess, a managing director and digital health lead for life sciences investment bank Outcome Capital, said a dizzying array of diagnostics and disease management and support tools have emerged, and the development of large-scale language models has further turbocharged the already rapid pace of AI tool development. According to him, physicians are using AI models daily to find better ways to communicate with patients, and the potential to apply the technology to diagnosis, patient monitoring, and testing over time is wide open.
"It'll be really amazing to see where some of this goes and where [the US Food and Drug Administration] will step in and regulate and provide some sort of guidelines," Hess said.
In the US, the FDA has already granted regulatory clearances for hundreds of devices that incorporate AI and machine learning-based software. In recent years, agency officials have been trying to give developers room to innovate and update their products already on the market through predetermined change control plans. In draft guidance, the agency said developers could outline in those plans the type and scope of changes they want to make without additional regulatory submissions, which would allow iterative improvements of machine-learning functions within FDA-reviewed limits.
Following pathology's lead
According to Volk, AI tools already have the power to help pathologists make better diagnostic decisions. A trained pathologist is still faster at reviewing tissues and identifying cellular maladies, but AI-trained software can help a pathologist find the atypical cells they would otherwise miss on a cytopathology slide containing 100,000 cells. It can also spot connections a pathologist might miss between those findings and clinical histories using "big data" to better serve patients, she said.
Eric Glassy, chair of CAP's information and technology leadership committee and a pathologist with the Affiliated Pathologists Medical Group, said on a recent media briefing that he uses image analysis software daily to help analyze immunohistochemistry results for breast cancer biomarkers or hormone receptors.
"I think the bottom line is pathologists need to make the most accurate and correct diagnosis they possibly can, and I use whatever tools are available to allow me to do that," he said.
He also predicted AI technologies will deliver long-term cost savings through efficiencies and precision medicine as payors see that pathologists using these technologies can perform fewer unnecessary tests and start therapy more quickly.
Matthew Hanna, vice chair of CAP's AI committee and a pathologist at Memorial Sloan Kettering Cancer Center, said in an interview that likely fewer than 5 percent of laboratories are using AI-developed tools in clinical practice to improve their accuracy and productivity or in research to help find biomarkers that could help guide treatments. About double to triple that number of labs likely use some type of image analysis software for the quantitative interpretation of digital slides, especially protein expression stains used to guide treatments, he estimated.
Most of the AI-based clinical tools so far are focused on oncology, such as to aid the detection of prostate cancer and determine the tumor grade, as well as to detect breast cancer and metastases in lymph nodes. By comparison, only a small number of AI tools have been developed for infectious disease identification.
Hanna said he uses AI to provide preliminary analysis of patient cases and to select which ones should be prioritized. If the software finds a patient has a tumor, he can begin examining those slide images and order any needed stains right away.
The AI software in his lab also provides visualizations, such as heat maps and opacity maps, to highlight areas that fit the model's criteria for tumors. Hanna declined to identify the developer or vendor of the software he uses.
"They help in finding small foci of tumor, they help in detecting metastatic breast cancer in the lymph node," he said.
Adoption curve
The increase of AI technologies in pathology has in part been driven by the uptake in labs of digital pathology technologies, some of which leverage AI.
Israel-based Ibex Medical Analytics said earlier this year that momentum was growing in the UK and US for its AI-based Galen digital pathology program. Among the firm's recent wins was a contract secured in March to provide its prostate and breast cancer diagnosis tools to 25 UK National Health Service Trusts. The company said that same month it had expanded an agreement with Alverno Laboratories for use of the Galen suite in Illinois and Indiana.
One of its competitors, Helsinki-based bioinformatics firm Aiforia Technologies, said it plans to build a suite of AI models and offer end-to-end clinical pathology workflow automation. Meanwhile, Paige gained US Food and Drug Administration de novo marketing authorization in 2021 for an AI-based pathology product, Paige Prostate, which is designed to help pathologists find small foci of cancer. Earlier this year it inked a deal for Microsoft to invest in the company's development of AI-based diagnostics and cloud-based patient care.
Adoption of those tools requires investment into digital pathology, whose adoption is rising despite hurdles, especially the high cost of adding digital pathology capabilities while the value and accuracy of image analysis algorithms are still being proven.
Hanna said most labs in the US have at least started the process of changing over from analog analysis to digitization of their glass slides, and the evidence he has seen indicates pathologists with digital imaging and machine learning will be able to screen samples more accurately and efficiently than those who don't. But he said unfamiliarity with these technologies and, among some pathologists, the fears they will be replaced by software is driving discomfort with adoption of AI/ML tools in pathology.
Cancer genomics firms such as Guardant Health also have been adopting AI technologies that can help automate PD-L1 scoring, while other firms in the space like Caris Life Sciences and Tempus have said they are eyeing the use of AI algorithms to hasten IHC interpretation, predict biomarkers, or guide pathologist review.
Diagnostics firms, meanwhile, are eyeing ways to use similar algorithm-based software to build out their product lines or use AI-driven pattern recognition software to bring together the data seen by physicians more quickly or effectively. Siemens Healthineers, for example, boasts on its website that it holds more than 700 patents related to machine learning and has established AI expertise and infrastructure that includes regional data centers and its Sherlock supercomputer.
Chuck Cooper, chief medical officer for diagnostics at Siemens, said it is also exploring ways to use AI models to discover earlier markers of sepsis and to identify laboratory errors, as well as to improve laboratory efficiency. Healthcare generates huge amounts of data that might be leveraged through these models to improve patient care or "even how the hospital system operates" through improvements to efficiency, effectiveness, and error detection.
"By using AI techniques, we can sort of coalesce larger sets of data that could generate insights that we otherwise might not see just with our own eyes," he said.
His colleague Dennis Gilbert, executive vice president and head of research, development, and innovation for diagnostics at Siemens, added the company is researching ways to use AI in product development including to design and test software in its products and to model prototype instruments.
AI models draw from the same set of biological information available to physicians, but the technology has use only if it can improve on the standard of care, he said.
"There are more biomarkers out there than we know what to do with," Gilbert said. "Taking a particular biomarker, attaching it to a clinical outcome, and proving that actually has benefit to the patient, that's the struggle."
As a result, Siemens is pacing its investments into AI to match the maturation of the technology. Cooper said the firm believes AI-backed technologies will offer significant value, but it is exercising patience on development because "it's going to take a lot of time and effort to understand and define what that value is and then position that properly within our strategy of how to generate value for our customers."
In the clinical space, a crowd of firms sees similar opportunities to integrate AI/ML-based tools into patient care.
BPGbio is collaborating with the US Department of Defense to validate a breast cancer diagnostic test that uses AI to analyze tissue and blood samples, while MeMed has developed a CE-marked and FDA-cleared test that uses AI algorithms to differentiate between bacterial and viral infections by measuring host immune response proteins TRAIL2, IP-103, and C-reactive protein.
Becton Dickinson secured FDA 510(k) clearance in May for an imaging application that uses AI to interpret methicillin-resistant Staphylococcus aureus (MRSA) bacterial growth, while Exact Sciences provides an AI-based platform to help clinicians interpret test results and develop patient-specific treatment plans.
Meantime, the US Department of Health and Human Services' Biomedical Advanced Research and Development Authority recently awarded a contract to machine-learning startup RAIsonance to develop software for cough-based disease detection.
Craig Steger, a director at Outcome Capital, said that beyond the previous investments he has seen into pathology and radiology, he is now seeing expanded interest in AI technologies and the use of large data signatures for oncology, cardiology, and neurology and simplifying the results to make diagnoses more palatable to patients.
"It is early days, but it really will impact, I think, how medicine is delivered, how we diagnose, but also how physicians talk to patients and interact with patients," he said.
Cowen analyst Dan Brennan noted that firms such as Adaptive Biotechnologies and Sophia Genetics are using AI and machine learning to differentiate themselves from the competition, but most diagnostics firms his team covers have been low-key about their use of AI technologies, quietly using those tools behind the scenes to analyze large data sets to find better signals for their tests.
The topics du jour instead are belt tightening in capital markets, cost cutting, and finding ways to grow. While he doubts firms are cutting AI programs, he said how much they are investing into AI and prioritizing its use depends on whether it is central to their business or improves their test performance.
Fellow Cowen analyst Steven Mah said the usefulness of these AI tools depends most on the quality of data used to train algorithms. Pathology is a natural fit because the field generates clear, objective data acquired through slide images and, similarly, he said the data generated through clinical chemistry also seem to fit well with AI tools.
"I think you'll see the early adopters more in those areas where you have really clean data and really clean traceability of that data," he said.
Mah is seeing more AI adoption in private companies rather than public ones, which tend to have greater cash conservation and capital runway needs. Those public biotech companies have instead been getting their feet wet by partnering with AI platform makers like Tempus and Caris on oncology applications and taking advantage of models that incorporate data from pathology, biomarker expression, and well-annotated clinical notes.
"Typically, with oncology, you have a wealth of data that's really well characterized and, oftentimes, it's multimodal," Mah said.
Barclays analyst Matt Miksic also wrote in a recent report that AI and algorithms have potential to improve patient outcomes and lower healthcare costs by helping clinicians identify and treat patients with undertreated or difficult-to-diagnose ailments. But he, too, cautioned that applying those technologies and securing reimbursement requires overcoming the challenge of generating evidence that the tools are safe, effective, and cost-effective.
He Sarina Yang, director of the toxicology laboratory at Weill Cornell Medicine, said in an email that machine-learning algorithms and AI-based tools have shown tremendous potential in an expanding array of uses in labs — such as identifying preanalytical errors, interpreting serum protein electrophoresis data, and improving lab test utilization — but the path toward implementing them into clinical practice remains long and laden with technical challenges, regulatory requirements, and the need to educate clinicians. Current barriers include privacy concerns, unintended bias in models, difficulties interpreting deep learning model outputs, susceptibility to irrelevant inputs, and a lack of reproducibility in some models.
"In addition, many laboratorians and clinicians are uncomfortable with the paucity of external validations as well as the lack of clinical accuracy expressed as sensitivity and specificity at various cutoff points for practical clinical application," she said. "It is appropriate for laboratories to consider ML models as test systems and evaluate these in silico tools in a manner similar to how in vitro assays are validated."
Fei Wang, director of the Institute of Artificial Intelligence for Digital Health at Weill Cornell Medicine, added that while machine-learning models are subject to bias through their training sets, end users are also susceptible to bias when they apply those tools and in how they interpret the results.
"We need a new generation of laboratory medicine informaticians who not only understand laboratory medicine specifics but also have the fundamental knowledge on AI and machine learning," he said.
Is there clinical value yet?
Siemens' Gilbert said the company has learned difficult lessons in recent years that the power of AI-based tools to find connections between markers is tempered by the need to put the data outputs into a clinical context. The field remains exciting, but he also recalled enthusiastic proclamations made in years past that genomic tests would predict patients' health needs for the course of their entire lives.
In August 2021, Siemens announced it had built a real-time AI-predictive tool, that uses nine laboratory parameters, to help clinicians identify which SARS-CoV-2 patients were at the highest risk of severe outcomes. Gilbert said identifying the connections between that profile and clinical outcomes was an exciting moment but creating that profile for a patient required combining various instruments with the work of multiple laboratory personnel, and the company struggled to turn that profile into a tool its customers could use and connect with the standard of care.
"We were all pretty excited and it was a great collaborative project with many thought leaders, but in terms of it having clinical impact, it really didn’t [have any] because we just couldn't get it to where it really needed to be to be effective," he said.
But the company remains committed long-term to using AI in developing its tools for customers. Cooper noted Siemens' instruments generate data across clinical chemistry, immunoassays, hematology, and hemostasis, presenting opportunities to figure out with the help of AI how to generate incremental value for customers.
"We're working very hard on some collaborations with leading academicians to figure out how we can do this sort of thing," he said. "This is not something we do in a silo. We need to work collaboratively externally and then also validate clinically in actual clinical settings."
The success of these tools will also depend on the ability to convince payors that the upfront cost of the technology will alleviate downstream costs, Outcome Capital's Steger said. In digital pathology, firms such as PathAI have partnered with large players in narrowly defined areas of oncology to show their impact on cancer patients, and that slow rollout and acceptance by oncologists, hospitals, and payors is going to help other firms see the path for adopting these technologies.
Steger said pathologists are retiring more quickly than new ones are entering the field, and tools that help a pathologist see more patients will potentially demonstrate value to hospital systems and pathology groups. Their questions are whether each technology will help a pathologist treat a patient more quickly or at lower cost.
"Can I see more patients?" he said. "Can I diagnose more, can I treat more, do I have better outcomes? Those are all the sort of metrics that they're going to be looking at."
Steger said he is excited to see the development in coming years of an AI model with continuous learning and growth, although such a tool would have a high hurdle to clear to secure regulatory approval.
The FDA has so far cleared more than 500 medical devices that incorporate AI or machine learning, according to Hess, who also noted that within the past six months, every firm seeking funding seems to have AI somewhere in their pitch deck.
He expects healthcare will someday incorporate technologies that include consultations with virtual physicians and AI-based preliminary symptom checkers. The diagnostic tests used by these models may need their own tests to maintain controls and accuracy over time.
"I think the future is wide open," Hess said. "I don't believe that most people have any clue what the next five or 10 years will bring, and, if they say they do, they're probably wrong."