The clinical trial process is broken for a variety of reasons, ranging from the difficulty of designing trial protocols to the cost of running them. But one critical (and solvable) issue is finding, enrolling, and retaining the right patients at the right point in time, efficiently. Fifty-five percent of all interventional trials terminated each year are terminated due to lack of patient enrollment.
These statistics paint a sobering reality. Potentially life-saving drugs often never become accessible to patients who have exhausted their chances with standard-of-care therapies. Physicians and healthcare systems lose resources, effort, and time spent on research that is ultimately terminated. Pharmaceutical companies lose investments made in a drug, while foregoing any potential revenue that drug could have produced. This administrative inefficiency is a malignancy that can and must be cured.
Part of the challenge of enrolling the right patients quickly is that trial enrollment criteria are often incredibly complex. For a single trial, a patient may need to meet 30 to 40 separate criteria. So, rather than simply looking for patients with breast cancer, a trial may be looking for patients with triple-negative breast cancer, a specific BRCA2 mutation, no metastases to the brain, but other types of metastases, a history of having tried a specific type of immunotherapy but not yet another—and other highly specific characteristics.
In some cases, these criteria are so strict that a sufficient number of eligible patients may never exist. More commonly, however, the challenge is temporal; while a patient may be a match one week, their condition may be rapidly changing, making them ineligible for the same trial just a few weeks later. This is especially common with cancer, which is, by its nature, a progressive condition. If a patient is not identified during the narrow window of time during which they qualify for the trial, they will not meet the criteria for participation.
Under the current system, physicians running the research at a hospital (principal investigators), either wait for other doctors to refer patients (which can take years–especially for rare conditions), or they have their team of clinical research coordinators performing manual review of patient charts for matches, which can be cost-prohibitive at $100 an hour and up to 30 minutes per chart (or electronic medical record). Moreover, by the time these clinical research coordinators have read some files in week one, those same files or remaining files may have changed by week two or three. It is not humanly possible for a researcher to read and process the data fast enough to find patients during their window of opportunity.
Unleashing the potential of AI
One promising solution is using a type of artificial intelligence (AI) called natural language processing to review clinicians’ notes via electronic health records to find patients eligible for clinical trial participation. It’s a process that takes seconds. But there’s a first step that’s critical before deploying the AI, which is streamlining the trial eligibility criteria so that the queries can be objective and deterministic in finding the exact data or relevant insights chronologically described in the physician’s notes.
This process of streamlining starts with deconstructing existing criteria into discrete data queries or “questions.” If a trial sponsor is looking for patients with metastatic, castration-resistant prostate cancer, for example, there are more than three data points the AI will be looking for in patient charts. One data point is metastasis status. Another is a diagnosis of prostate cancer. Castration-resistance is a bit more complicated. If a clinician has not used these terms specifically in their notes, the AI will need to be programmed to algorithmically “know” that increasing PSA levels while testosterone levels are decreasing indicates a sign of castration-resistant disease. Subsequently, it will need to then find multiple data points showing the patient’s PSA and testosterone levels over time.
In essence, protocol scoping means turning a complex protocol into simple computer-friendly queries, similar to taking the chemical composition of food and turning it into a recipe for a specific dish.
The process described requires sophisticated Natural Language Understanding (NLU) capabilities. Historically, NLU for clinical notes has been synonymous with Named Entity Recognition (NER), a shallow technique that simply identifies entities (e.g., medications, diseases, diagnostic procedures) in the text of a clinical note. However, NER does not provide the sort of structured relational semantics that are necessary for a genuine understanding of clinical notes. It might identify a mention of a drug name such as prednisone, but it will not tell you what the clinical note is saying about prednisone (e.g., whether it was started, stopped, resumed, or whether and why the dosage was increased or decreased). True NLU capabilities are needed to generate nuanced insights from extensive notes in medical records.
Increasing comfort with AI among clinicians and the public
AI is still a work in progress, but it’s helpful to remember how quickly the technology is advancing. It was only a few years ago when we were asking Siri the weather, and now, we’re using AI to review unstructured data in a clinician’s notes with near human accuracy for many clinical characteristics.
The ultimate goal is to improve treatment options for patients—and extend lives. Take a condition like uveal melanoma with metastasis to the liver, a rare form of cancer that has an average life expectancy of 16 months after diagnosis. Since there are no FDA-approved treatment options for the disease, the current standard of care is to refer eligible patients for clinical trials. As of March of this year, only 11 clinical trials had been initiated for this condition since 2011 in the US; three of those were terminated, and in total, fewer than 400 patients have been enrolled in all the clinical trials. Without access to a clinical trial, these patients only have a 6% likelihood of survival. So it’s not an overstatement to say that by accelerating the recruitment process and improving the efficiency of clinical trials, AI can save lives.
Despite the promise of AI in the clinical trial enrollment process, overcoming clinician resistance remains an obstacle. There is fear among some doctors that AI seeks to replace them. But AI-driven tools shouldn’t interact with patients, or even know their identities. Their goal should be simply to review de-identified and raw medical data to find patients who are eligible for active clinical trials. The conversation about treatment options, trial participation options, and other care plan decisions should still always between the clinician and the patient.
I’m hopeful that as clinicians and trial sponsors see what cutting-edge AI tools can do, they’ll become more comfortable with the capabilities of new technologies. By combining AI with human expertise and oversight, we have the potential to revolutionize the clinical trial process—and most importantly, extend and save lives.
Photo: Deidre Blackman, Getty Images