If the requirements are met, an effective AI-enhanced prescreening tool will allow more principal investigators to run clinical trials and more patients to gain access to potentially lifesaving treatments.
Prescreening potential candidates is typically one of the most significant limiting factors in running oncology clinical trials. Identifying eligible participants requires manual effort and resources from trained medical professionals to flag prospects, scan their charts and confirm eligibility. The process creates delays, increases costs and burdens trials that cannot find enough appropriate patients to evaluate results.
While there has been significant progress in developing artificial intelligence tools for pulling information out of large data sets, no automated method for using an AI to assist in clinical trial prescreening has yet demonstrated its usefulness.
The reasons for the failure lie not only in the difficulty of developing the sophisticated algorithms required to deal effectively with unstructured clinical oncology data, but also in integrating the clinical AI into the procedures, practices, data flows and administrative requirements of a busy clinical setting.
The increasing problem of patient prescreening
Identifying patients who might be appropriate for a specific oncology clinical trial requires understanding the disease state of a large number of potential patients to ensure that they meet increasingly complex and specific eligibility criteria.
Electronic medical records (EMRs) were originally designed to facilitate billing, and to document provider assessments, activities and treatment plans. Providing structured diagnostic data for this kind of analysis was not its primary purpose.
However, somewhere between 80 and 90 percent of medical data is unstructured, useful only when read by humans, not by machine. This data consists largely of scanned documents, along with such things as pathology reports, lab reports and clinical images.
The non-machine-readable nature of this data — coupled with the fact that the most useful clinical data resides in unstructured notes and documents — means that patient prescreening is still a highly manual process. The prescreening is typically conducted by a traditional chart review by a trained medical professional who scrutinizes each prospective subject’s medical record, extracting relevant matching attributes, and often relying heavily on physician notes flagging a potential match during the course of regular treatment.
This prescreening process is slow, costly, misses possible eligible patients and imposes significant challenges on already burdened clinics and hospitals.
There has been real optimism about using AI and natural language programming to automate much of the prescreening process. Unfortunately, the promise has not yet been fulfilled, largely because of insufficient attention to the complexity of the clinical system inside which the AI must work.
What has kept an AI solution from working until now
There are five specific factors essential for an effective AI-supported prescreening process. Some are obvious, some get neglected and some can be surprisingly difficult to accomplish.
1. Data extraction
At the start, the AI needs to be able to extract information from the mass of unstructured clinical data in the medical records. This is fundamental and has been the focus of a lot of recent research, but extraction itself is not understanding.
2. Natural Language Programming
Once the information has been extracted and de-identified, it needs to be fed into an oncology-specific natural language processing engine to identify the information in the record that is pertinent for trial matching. This is a hard problem because clinical language is, at least, just as vague, ambiguous and implicit as ordinary language.
This means the AI must be specific to a therapeutic area, such as oncology, in order to continuously train and improve the models. The more general the AI, the poorer its performance will be when faced with specific tasks of even moderate complexity. Only domain-specific AIs can perform well on this type of specialized data.
3. Guidance by clinical staff
It is also important to understand that even a highly capable AI is still a tool, with the purpose of augmenting human abilities, not replacing them. Only clinicians experienced in oncology can refine and guide AI recommendations.
4. Access to large clinical data sets
To continually improve, the AI needs large amounts of patient data, from as many heterogeneous sources as possible. As anyone who works with healthcare data knows this is annoyingly difficult to acquire and make useful. Healthcare data is siloed, stored in a wide variety formats and sometimes difficult-to-access locations, and guarded by a wide range of access rules.
Anyone hoping to have an AI useful for oncology clinical trial prescreening will need to have relationships with a large number of hospitals and clinics, and experience in securely acquiring, transferring, storing, and processing clinical data — reliably and at scale.
5. Process and contract integration
To become useful, clinical AI must become part of the operations of the hospital or clinic it serves. Like any new technology relationship, it will require security reviews, IP assessments, a slot in the budget cycle and to go through the entire IT onboarding process, which can take months. Ideally, whoever is providing the AI already has an intimate relationship with IT, contracting, finance and all of the other stakeholders.
The opportunity is real
So, while industry is exploding in its investment and creation of AI, it is essential to know that in order to be utilized, AI must still follow all of the rules that apply to any of the other increasingly complex pieces of software and hardware that are a major part of modern healthcare. Neglecting these requirements will lead to a failed attempt to improve the oncology clinical trial prescreening process, already disliked by physicians and avoided by sponsors.
But if the requirements are met, an effective AI-enhanced prescreening tool will allow more principal investigators to run clinical trials, including those at smaller sites serving a variety of populations, and more patients to gain access to potentially lifesaving treatments.
Date: June 11, 2019