NEW YORK – Through a partnership with artificial intelligence firm OpenAI, Color Health has developed software for analyzing patient medical records and helping physicians create personalized cancer testing plans.
The Burlingame, California-based firm intends to use the tool to aid testing decisions for hundreds of thousands of patients by the end of this year.
The company said recently that it had used OpenAI's GPT-4o system to develop its Copilot software for healthcare providers to help clinicians identify which screening tests could help them find early-stage cancers while patients have a higher likelihood of survival with treatment. Additionally, Copilot may accelerate the start of treatment following a cancer diagnosis by determining what other test results may still needed for an oncologist to develop a treatment plan. Copilot has been able to provide results in about five minutes.
Although Color began as a genetic testing firm, it has shifted in recent years to commercialize cancer detection and care programs that are offered to the largest healthcare purchasers such as employers, pensions, and unions.
Color CEO Othman Laraki said that the company's focus is now on improving patient outcomes through adherence to screening guidelines, earlier diagnosis, and value-based care. To that end, the firm has partnered in recent years with the American Cancer Society to offer comprehensive cancer detection services through employers and unions and provide support services for cancer patients and survivors.
Laraki said that its collaboration with OpenAI on Copilot, meanwhile, has been focused on the development of an AI-based tool to interpret patient data and analyze dense healthcare guidelines. Such a tool could help to improve cancer outcomes through more consistent adherence by the general population to risk-adjusted cancer screening guidelines as well as the creation of pre-treatment testing workups ahead of a patient's first visit with an oncologist following a cancer diagnosis.
Copilot analyzes a patient's medical record including previous lab test results, biopsy results, and family history of cancer and make guidance-based recommendations. The firm said on its website that doctors already have access to cancer testing guidance through myriad sources such as the American Cancer Society, National Comprehensive Cancer Network, and US Preventive Services Task Force, as well as through publications such as the Journal of Clinical Oncology, and the Lancet, but keeping up with changing guidelines and recent clinical trials can be challenging whereas large language models can shine a light on the most relevant information.
Laraki said that Color has seen through its history as a cancer testing company that about 30 percent of people have some non-standard factors that should be considered when developing their cancer screening schedules including hereditary risk, smoking history, occupational exposures, lifestyle and socioeconomic risks, as well as broader demographic risk factors.
Cancer screening guidelines are often poorly implemented and clinicians can miss opportunities to catch cancers earlier, Laraki said. Primary care physicians also typically spend little time with patients and often have not memorized or implemented all of the risk-adjusted guidelines.
"You take an average PCP, it tends to be pretty difficult for them to manage a woman with a very high degree of ovarian cancer risk, for example, when they don't necessarily know what are the risk-adjusted guidelines," he said.
While the US Preventive Services Task Force recommends, for example, that many adults with a 20-year history of pack-a-day smoking receive annual low-dose CT screening, Laraki said that only a tiny minority of those patients receive that screening even as adherence to recommended screening has been shown to help catch and treat cancers earlier, improve survival, and decrease healthcare costs.
Laraki said that the company is also developing Copilot to analyze patient records and test results following a cancer diagnosis and create a pretreatment workup ahead of the patient's first visit with an oncologist. He said that patients often wait for weeks following a diagnosis for their first visit with an oncologist and their treatment is further delayed while they wait for the results of tests that are ordered for the pretreatment workup.
"Every day or every week that you delay initiating treatment has a pretty big effect on survival rates and the cost of treatment," he said.
Based on information gleaned from Copilot, primary care providers or oncologist can order those tests immediately following diagnosis and begin treatment weeks to months earlier. The model would identify the information that would be needed by an oncologist based on established guidelines as well as the patient's medical record including their medical history and the results of previous tests and imaging.
Laraki said that the healthcare providers who use Copilot will remain involved in any testing decisions. The doctors who have used the tool so far to create patient-specific pretreatment workups have said that it helped them to develop more comprehensive pretreatment testing plans with greater efficiency, he noted.
Copilot also provides references and resources with its recommendations so that healthcare providers can check the results and correct any problems.
Matthew Schabath, program leader for the cancer epidemiology program at Moffitt Cancer Center, said that he has worked on the development of personalized, precision risk-based modeling and he thinks that models that account for a multitude of patient-specific parameters could improve patient risk assessment. Such models, however, are difficult to develop and they are subject to limitations including the limits of knowledge of why many people develop cancers in the absence of known risk factors.
Schabath said that, before using any tool used to guide cancer testing, he would want details on what data are being used for the assessments and how such an AI-based model was trained, tested, and validated. That would include how the company accounted for fairness and bias, whether the results are generalizable, and whether the work has been vetted through peer-review.
Other experts recently said that they, too, see the potential that rising use of AI and machine learning-based medical devices could either help to improve equity in healthcare or exacerbate bias and inequities. At the same time, the developers and regulators of AI-based technologies are still figuring out how to deal with a changing landscape that could include innovative models with continuous updates as opposed to stable algorithms.
In a statement Color Health said that it has validated its model with cancer experts as well as worked with clinical experts and advisors to test the model's performance and compared the results against expert recommendations. The company noted that it developed a knowledgebase of clinical guidelines and standards of practice and codified them so that they could be understood and analyzed by their Copilot model. Because GPT-4o is a pre-trained generalist large language model, the company was able to focus its work on curating validation and test sets.
Laraki said that the company is partnering with the University of California San Francisco on the development of the pretreatment workup component of Copilot. OpenAI recently said that the partnership with the UCSF Helen Diller Family Comprehensive Cancer Center would involve a retrospective evaluation of the software's performance followed by a targeted rollout and, potentially, integration of Copilot into clinical workflows for all new cancer cases.
That research will involve the analysis of patient records to make sure that the pretreatment workups have high concordance with orders from oncologists and that use of Copilot is connected with faster times to begin cancer treatment.
Color Health has been rolling out the Copilot tool for cancer screening through select physicians and Laraki said that the company has set a goal of developing screening plans for about 200,000 patients by the end of 2024. The firm is focused on implementing those screening tools through its programs in partnership with the American Cancer Society. He said that Color also plans to begin offering its pretreatment workup application for certain groups of cancer patients at UCSF starting in Q4 of 2024, although he said that the company is still deciding which cancers to target first for pretreatment workups.