By Deborah Borfitz
October 21, 2020 | Surging use of telemedicine technology since the COVID-19 pandemic hit the U.S. earlier this year has resulted in a massive influx of unstructured patient data to sort through when searching for drug safety signals and compiling adverse event reports required by federal regulators. Physicians are understandably more scripted when interacting with patients via a video screen rather than in person, as well as more “expressive” in their visit notes, because signs of trouble can be subtle when viewed remotely, says Updesh Dosanjh, practice leader for technology solutions at IQVIA.
Patients who complain of a pain in their leg can’t as easily show their doctor where and how it hurts via telemedicine than in a traditional office visit, for instance. Providers might also miss breathing problems that would be visually obvious during a live exam.
Natural language processing (NLP) thus has a potentially bigger role to play in automating the pharmacovigilance process, starting with interpreting data in free text fields and doing some of the cumbersome analytical work, Dosanjh says. This would include searches for relevant clinical cases in the medical literature and key words and themes on websites and social media channels.
Only about 10% of adverse events ever get reported to a doctor—irrespective of the visit type—but may instead be informally logged online via chats, tweets, and Facebook posts in a multitude of languages, notes Dosanjh. NLP is a good fit for that level of detective work. If the out-of-pocket cost for a drug or drug delivery device is negligible—which is more the case outside of the U.S.—people are unlikely to go through the bureaucratic hassle of self-reporting on product issues (e.g., a malfunctioning EpiPen or a drug that causes a mild rash or nausea) to the manufacturer.
Trolling the internet for clues is an exercise no human could tackle with any degree of success. For example, a review of 600,000 web posts that IQVIA’s Vigilance Platform did for one client only generated 20 suspected adverse events linked to a portion of its product portfolio, a percentage so small it would have likely been lost to human error, Dosanjh says.
NLP can be similarly useful in ferreting out important contextual information for clinical trials. Cannabidiol (CBD) therapy studies, for example, might be supplemented with real-world cannabis side effects data that would be difficult to capture in a controlled scenario, says Dosanjh.
“I’m not sure that most [CBD] companies even know that this is possible, but there’s certainly a treasure trove of information available to them should they want to go for it,” he says. Outside of CBD, trial sponsors might instead want to deploy NLP to semantically analyze what people are communicating to their friends and family about certain drugs that could augment standard data collection efforts.
Clinical trial sponsors in general have become “a bit more proactive” about preventing study participants from experiencing adverse events, Dosanjh adds. One way to identify potential problems when conducting trials is to boost the number of study visits but do them virtually to reduce the time and cost burden, and currently, COVID-19 exposure risk, on participants. This will generate increasingly large volumes of data for NLP to process, allowing companies to do more with the same resources.
Inversely, this ability to automate trials reduces the cost of studies by enabling them to be more effective even with smaller patient populations, he says. In this case, leveraging NLP's ability to read “between the lines” empowers companies to find patterns within the data but with less pieces of information than human analysis.
Getting Smarter
Traditionally, drug safety signals have been detected retrospectively via statistical analysis of data collected over several months, says Dosanjh. It was also hoped that physicians were capturing and responding to concerning adverse events or safety issues as they were seen first-hand.
The added value of NLP is, at minimum, to “oversee what physicians are interpreting out of the data… in real time,” he continues. It could be running in the background and, as needed, prompting doctors to consider something they might have missed or could be explained in a different way.
Using the same Vigilance Platform from IQVIA, NLP could also streamline the completion of periodic risk assessment from the deluge of unstructured data while bringing in relevant cases or narratives from cases to expedite the analysis work. The end result, Dosanjh says, is a more meaningful, contextualized view of the information than would otherwise be available with traditional monitoring and more manual reporting.
Unaided by NLP, interim analyses have proven to be correct less than 30% of the time among client companies, he says. “Every single case that they’ve entered had some errors in it.”
What people often forget is that NLP gets “smarter” every time it is used, Dosanjh adds. While it has already solved the huge problem of normalizing data cycles, that does not encapsulate the full extent of what the tool can do. The ideal scenario is to have it “listening and learning” from every interaction and part of a case at once, from start to finish.
It should soon be technologically possible for NLP to listen in even on video visits with patients and, when doctors forget to report an adverse event, do it for them, continues Dosanjh. NLP is already doing this sort of “voice-to-text simulation” work in call centers and throughout the bank and airline industries, using chatbots to remind workers in real time to ask pertinent questions.
“A combination of audio extraction and audio analysis, or video extraction and video analysis, can be used to create a voice-to-text transcription and look for any unreported adverse events,” says Dosanjh. “However, this method is still in its early stages in terms of improvement and depends heavily on the languages.” Consequently, within pharmacovigilance, it’s currently being used as a quality check tool to ensure that no adverse events went unreported. “Following its use, a human will need to review, verify, and extract the data.”
This added layer of monitoring helps to reduce the likelihood of missed patterns and reporting errors, he explains, noting that NLP for pharmacovigilance is still a “relatively novel approach.” NLP’s capabilities, however, can be taken even further.
Feeling Heard
The use cases in play currently include a combination of optical character recognition (OCR) and NLP to extract the relevant case information needed to reliably report adverse events—about half of which come in via email—as well as to analyze images of adverse events with human follow-up to catch any errors, he says. OCR and NLP can also be used to look for unreported events within unstructured data (e.g., in electronic health records) and use signal detection to dependably identify any adverse events.
The Vigilance Platform is most often used to turn incoming adverse events, in whatever format they originally exist, into a structured field in the safety database of a life science company, Dosanjh says. The task involves translation, since up to 40% of all information received is not in the English language. When it comes to languages such as Spanish, German, French and Italian, the software is “as good as if not better” than human translators.
Next-generation NLP will give physicians search results before they ask for them, the same way Google anticipates the type of information users are seeking and patients need, Dosanjh continues. “If NLP can detect the inference or feelings of patients, which it can kind of do, maybe it [suggests] that they be given one version of a [medicine information] leaflet or another” and improves the likelihood that the drug will be taken as intended.
In addition to potentially improving medication adherence and safety, NLP might also help patients feel more listened to and understood—a chronic shortcoming in their current interactions with medical professionals, Dosanjh says. Similarly, in a clinical trial setting where NLP is transcribing data and detecting adverse events, investigators could “just focus on talking to patients... significantly increasing the odds that they’ll stay in that trial,” reducing costs for sponsors while also improving patient outcomes.