functionbar_help
Font Size
functionbar_contact
Aboutus
Search Results for ""
Share
Top Stories

Published On September 22, 2009

CLINICAL RESEARCH

Between the Lines

Patients on networking sites discuss their illnesses and treatments. Can pattern-recognition software pull insights from the noise?

EVERY DAY, THOUSANDS OF DEVOTEES LOG ON to the patient community Inspire.com. They check up on one another, ask for advice and tell stories of their illnesses. Though most have never met in person, they’ve come to know one another well, and they often share very specific experiences. “Day 5 of Tarceva, and I noticed about 4 pink welts on the back of my tongue,” writes a woman identified as Sandy45 in a post to a cancer group’s discussion forum. “Actually my entire mouth burns when I drink soda. Doesn’t look like any of the pictures I have looked up.”

A woman who goes by the handle Elaineb2 writes back, “I saw my oncologist Thursday and this is just one of the symptoms of Tarceva. I also have peeling fingers on both hands and my Onc. is real pleased as this (as well as the rash) shows that the Tarceva is doing its job. Just take one day at a time and TRY not to worry. LOL,” she concludes with the online shorthand for “laughing out loud.”

For Sandy45 and Elaineb2, both coping with advanced lung cancer, trading information about treatments and startling symptoms has become part of a reassuring routine. What they learn from each other, and from other members of their online group, may go beyond what they’re able to find through other sources. If they checked the safety information on the manufacturer’s site for Tarceva (erlotinib), for example, they’d learn only that “bullous, blistering and exfoliative skin conditions have been reported, including cases suggestive of Stevens-Johnson syndrome/toxic epidermal necrolysis, which in some cases were fatal.” Nowhere on the official site would they discover the notion, relayed from Elaineb2’s oncologist, that such side effects in a mild form might be a welcome sign.

Members of Inspire take part in discussion groups on hundreds of diseases and conditions, and it’s obvious how much participants depend on one another for information, inspiration and support. But all of their online “talk,” which generates some quarter-million words a day, may have another use as well. It’s a giant pool of glimpses into patients’ everyday lives—glimpses that aren’t gathered with the tools traditionally employed by researchers, who conduct studies according to strict observational protocols, with painstakingly calibrated surveys. Still, some scientists have begun to explore ways to mine this ever-deepening vein of online content.

By using language- and pattern-analyzing software to search and distill the cascade of words on social networking sites, researchers may be able to identify trends and signals, such as repeated mentions of a certain drug in connection with a particular side effect or treatment success. “Listening” to thousands of patients as they discuss treatments, side effects and experiences with their diseases might even help scientists come up with hypotheses worthy of study. “There’s a tremendous amount that goes on in a patient’s daily life,” says Frank Moss, director of the Massachusetts Institute of Technology Media Lab, an emerging-technologies research group that helped create one disease-related networking site, and a founder of the cancer drug company Infinity Pharmaceuticals. “They’re trying off-label drugs, different diets, different exercise, different lifestyles. That information isn’t easily available to clinicians, but the cure to the disease may lie within it.”

IN THE DAYS BEFORE THE INTERNET, patients had few ways to learn more about what ailed them. But the advent of the Web meant they could look up their symptoms, read descriptions of their conditions on consumer medical sites and even delve into published medical literature. They no longer had to take their physicians’ assessments and advice on faith. And perhaps most important, they could identify others who had the same medical problems.

Most early online disease communities revolved around e-mail lists, and many groups that pioneered their use still play important roles in organizing patients and disseminating information. The Association of Cancer Online Resources, for example, was founded in 1995 by Gilles Frydman after he ventured onto the Internet for a second opinion about his wife’s breast cancer treatment options. ACOR now coordinates 159 mailing lists that deliver 1.5 million e-mails a week.

What makes Inspire and other so-called social networking services different is a set of programming techniques and computer tools that make it simple to establish interactive, community-based Websites. Every day more than 120 million people spend time on Facebook, the most popular social networking site. Facebook members are able to recruit “friends,” view posted photographs and videos, share Web links they think might be of interest and write on one another’s “walls” in public exchanges. They also post profiles that define their online identity—their schools or jobs, favorite movies, pet peeves. The site’s technology is constantly evolving, providing new ways for people to stay in touch with one another and with an expanding web of groups of people with similar interests.

Adapted for medical networking, the Facebook model gives participants easy access to information and the chance to contribute their own observations. Consider a woman diagnosed with sarcoidosis, an autoimmune disease of unknown origin, who finds her way to Inspire’s Stop Sarcoidosis Support Community. Joining involves simply providing an e-mail address, choosing a screen name and establishing a profile that may include a photograph and biographical details. The site offers links to basic information about the disease and current research, lists events (a fund-raising golf tournament in Denver, a “hike for lung health” in Chicago) and provides details about local support groups. Members are invited to start a discussion, post a journal entry or “meet others like you.” Each morning, members receive an e-mail highlighting the site’s latest discussions and journal entries.

Those personal posts are at the heart of the sarcoidosis group and typical of the conversations about diseases on Inspire and other social networking sites. One 40-year-old woman from Georgia, identified as denlock, asks in a recent post if anyone else has “ever felt light headed or kind of dizzy. Not just as a spell but all the time.” She writes, “Sometimes I have problems breathing but some of my granulomas have shrunken in my chest cavity but not the 2 larger ones on my lungs. I should tell someone but I don’t know how...how can you diagnose dizzyness?”

This poignant uncertainty draws immediate worry and comment from other posters, who describe at length their own histories with dizziness. They tell denlock to talk with her doctor, which she agrees to do. One patient picks up on her use of prednisone, which denlock had assumed was irrelevant, but in this poster’s case had caused her blood sugar to fall dangerously. “I will also make a note of this and bring to the doctor’s attention,” writes denlock. “Maybe I just need a full workup of bloodwork to make sure everything is ok.” The discussion continues.

MINING THAT KIND OF BACK-AND-FORTH for medical insights wasn’t part of the purpose when Brian Loew and two colleagues founded Inspire in 2005. Rather, the site’s business plan was based on helping people find and participate in clinical trials. Human studies face chronic shortages of participants; meanwhile, many desperately ill patients would gladly participate if only they had the opportunity. Inspire works with dozens of disease research groups, such as the Lung Cancer Alliance and the National Osteoporosis Foundation, whose disease-specific groups form part of the site. It also receives funding from four pharmaceutical companies that pay a flat fee for Inspire to recruit from its 125,000 members. Members learn of trials through alerts targeted to their condition, treatment, age, gender and geographical proximity to test sites. One Inspire-organized lung cancer trial has already been completed; another, for an arthritis drug, is under way.

The idea of probing Inspire members’ conversations came later, with help from Simetric, a marketing company that specializes in social media analysis. Inspire and Simetric took an approach called natural language processing, in which a computer program scans informal conversation for key words and their context. As long as a sentence contains some sort of sentiment, the software will notice; it figures out whether the author’s feelings were positive or negative, and it links those feelings to a treatment or symptom under discussion.

A query for mentions of the multiple sclerosis drug Avonex, for example, would parse a post by one user who writes, “i felt worse on avonex than my ms made me feel. while on avonex my psoriasis got VERY VERY bad/worse.” Another post reads, “I have been losing my hair…. I am going to switch from AVONEX to COPAXONE…to see if it is the AVONEX that is causing my hair issues.” After the program sorts through the text, Simetric employees review the results, double-checking the computer’s interpretations and dealing with tricky cases. The software may, for example, have trouble with the apparent contradiction of “wicked good” or pass over phrases it hasn’t been programmed to recognize.

A potentially significant use of software-driven analysis would be to study the effect of cancer drugs on quality of life. “One thing women with metastatic breast cancer discuss is the notion that quality of life is not a significant enough part of treatment,” says Loew. “Doctors talk about extending life by so many months, but they don’t talk about what life will be like during those months.”

Some quality-of-life metrics—for mood, anxiety, coping—have been developed for mental health research, says Stewart Fleishman, director of cancer supportive services at New York City’s Beth Israel Medical Center and St. Luke’s–Roosevelt Hospital Center and vice chair of a quality-of-life committee within a clinical trials group sponsored by the National Cancer Institute. Applying those standards to cancer drug trials can require thorough rewriting. For example, it probably doesn’t make sense to ask cancer patients whether a drug tells them to do things. Even basic side effects, such as low-level nausea, can be difficult to quantify.

Discussion analysis could provide an alternate way to assess quality-of-life issues. It might also help in evaluating drugs after they have been approved. Social networking conversations about a particular drug may contain information about effects that weren’t detected during comparatively brief Food and Drug Administration trials. “The side effects of a drug may be tenable if you’re taking it for a short time but have a profound negative impact on quality of life if you take it for years,” says ACOR’s Frydman.

WHILE INSPIRE IS LOOKING FOR LARGE-SCALE PATTERNS—such as the connection between Avonex and hair loss or flulike symptoms—a somewhat different approach is being taken by Ian Eslick, a doctoral student at the MIT Media Lab, which helped create LAMsight. The networking site is an offshoot of the LAM Treatment Alliance, an organization driving treatment research for people with lymphangioleiomyomatosis, a rare, incurable lung disease. Discussion is just one way LAMsight’s members generate information. They also use the site to enter detailed accounts of their physiology, symptoms and treatment regimens. They’re invited to participate in online surveys about the possible effects of alternative or complementary medical approaches; the prevalence of incontinence as a LAM symptom and experiences in managing it; reproductive health histories; and other topics related to living with LAM. All this information will eventually provide fodder for further research by clinicians and academics, and even by the patients themselves.

Eslick is working on the development of a language analysis tool that can pick out potentially interesting pieces of conversation—posts that differ from the norm or a smattering of similar reports. “It’s not so much sentiment analysis as figuring out the language of a particular patient domain,” he says. These patterns may reveal potential relationships—for example, between a low-fat diet and a slowing of symptom progression—that researchers could then study in a more targeted fashion.

“Patients know their diseases best, and the benefit of mining patient forums is to pull out highlights of what they know and to generate hypotheses,” says Eslick. “You use patient knowledge to guide the way you use observational data.” If there’s additional information that might be relevant, LAMsight users can add it to the list of measurements they enter into the site each day.

Much of this is information that academic researchers and drug developers might otherwise never consider. A large percentage of work on treating or curing diseases involves genetic studies and laboratory experiments seeking molecular targets. But if, for example, analysis of LAMsight data does suggest a link between low-fat diets and a slowing of disease progression, further studies could be designed to explore that possible connection in a more controlled way, with professional monitoring of symptoms and diets. “Ideally, analyzing the kind of data our site generates would lead to strongly supported hypotheses about disease that will prompt new research,” says Eslick. “We’ll find repeatable phenomena that won’t occur to researchers who are focused on biological mechanisms and looking for drug targets.”

Patients linked by social networking sites constantly compare their experiences of diseases and possible treatments, and word of unusual successes spreads quickly. Eslick was inspired by a story related to him by George Demetri, director of the Center for Sarcoma and Bone Oncology at Dana-Farber Cancer Institute and an adviser to LAMsight. A patient suffering from a rare blood disease had heard of positive results in people taking Gleevec, a leukemia drug, and asked his doctor to try it. There was no reason to think it would work, but the patient responded powerfully.

“The clinical discovery preceded the science,” wrote Demetri in an e-mail. Researchers went on to find the molecular mechanism by which Gleevec targeted a previously unknown mutation in abnormal blood cells. Added Demetri: “How many other unidentified pathways might there be?”

INSPIRE WILL EVENTUALLY MAKE ITS ANALYSIS TOOL available to outside researchers and to site members. LAMsight is also providing analysis tools to users—not just discussion analysis but also programs for interpreting other types of data. “It gets interesting when you say, ‘How do you let patients take their own information and turn it into a causal graph?’” says MIT’s Moss. “You provide tools to patients that let them mine their intuitions. We structure their conversation and enable them to come up with useful hypotheses.”

These are still very early days for mining patient networking sites, and both Loew and Eslick caution that their methods are being refined. Other researchers echo their warnings. “The analysis of patient conversations will itself have to be subjected to scientific evaluation,” says Patricia Ganz, a cancer treatment specialist at the University of California at Los Angeles. “Until then we won’t know whether they add value.”

Among the concerns are whether a few especially vocal members might skew discussion content, which can sometimes move into the realm of rumor. The reliability of data could also be affected if people are too suggestible—if, for example, they begin to notice a particular side effect of a treatment only after someone else has talked about it. And unlike in a formal study, in which variables are carefully controlled, drug effects mentioned on a social networking site might relate to information about a patient that doesn’t show up in the data.

Yet even if the information is flawed, analyzing it may suggest possibilities that researchers would not otherwise have considered. “The key is hypothesis generation and exploratory data analysis,” says Dan Lizotte, a University of Michigan biomedical statistician. “It would need to be followed by rigorous verification. But it would be insane not to use this data.”

 

DOSSIER

1. “Learning From E-Patients at Massachusetts General Hospital,” by John Lester et al., British Medical Journal, May 15, 2004. A summary of how people behave and what they share in online medical forums, and an early recognition of Internet communities by the medical establishment.

2. “PatientsLikeMe: Consumer Health Vocabulary as a Folksonomy,” by Catherine Smith and Paul Wicks, AMIA Annual Symposium Proceedings, November 2008. An overview of how members of online patient community PatientsLikeMe talk about disease, the article describes how their informal language structures can be classified—a necessary first step in analyzing them.

Share