Posted by Rob Knies

Distribution of medical queries per concern type

A recent study by the Pew Research Center indicates that 80 percent of adults in the United States have searched for medical information online. Such a figure underscores the fundamental importance that humans place on their health and wellbeing—and their reliance on the web for health information.

Microsoft researchers, however, have studied challenges with the use of the web for performing diagnosis and have pursued an understanding about why searches about common symptoms can escalate quickly into users focusing their attention on rare, serious conditions.

It’s called cyberchondria—making the leap from prosaic symptoms that could be explained in a multitude of ways to potential ailments more worrisome or debilitating.

Ryen W. White and Eric Horvitz have been studying this subject for several years and have just published their latest related  paper, Studies of the Onset and Persistence of Medical Concerns in Search Logs, to be presented Aug. 13 during the 35th annual conference hosted by the Association for Computing Machinery’s Special Interest Group on Information Retrieval (SIGIR 2012), which continues in Portland, Ore., through Aug. 16.

The authors seek to develop methods and tools to improve the search experience for those seeking health-related information. In their work being presented during SIGIR, they describe a longitudinal, log-based study of medical search and browsing aimed at determining how medical concerns emerge, persist, and influence behavior over time. Based on data from 170,000 consenting toolbar users over a three-month span, the study is the first of its kind.

“The methods described in this particular paper,” says White, a senior researcher in the Context, Learning, and User Experience for Search group at Microsoft Research Redmond, “allow us to learn health search preferences and processes to facilitate more timely and accurate search-engine support.”

White and Horvitz employed machine learning to construct probabilistic models that predict transitions from searches on symptoms to searches on conditions—and the escalation of web users’ medical concerns to make the leap to consideration of serious illnesses.

The latest paper studies the influence that prior concerns have on future behavior. Previous studies have focused on medical search behavior in single sessions, but this one examines medical searches spanning multiple sessions connected over time.

“The study focused on long-term medical search, with a focus on the onset of conditions queries in searchers’ long-term medical search behavior,” White explains. “We align users on the first occurrence of conditions in anonymized Bing logs and study their pre-onset behavior. We explored features of the pre-onset search behavior and investigated trends in the feature values and changes in their intensity over the pre-onset phase."

The research, he says, suggests several refinements that could help web users avoid anxiety while searching for health-related content, including:

  • Once a user begins searching for a condition, previous search clicks on related symptoms and conditions could be used as implicit feedback to augment current search results.
  • If a search engine can predict the onset of a condition search, previously sought symptoms could lead to estimation of a condition search to come and adjust the search experience based, for example, on severity or prevalence.
  • Prediction of between-session escalation could help inform the user before actual escalation, perhaps by offering a set of common, benign explanations.

“The web is an unprecedented resource for health care,” says Horvitz, a distinguished scientist at Microsoft Research Redmond. “However, we’ve identified problems with the popular use of the web as a diagnostic system—arising at the intersection of human psychology and information retrieval. We’ve been working to characterize and address the problems to enhance access to health-care information.”

This work has been featured in The New York Times, and the researchers’ previous papers include the first examination of the cyberchondria phenomenon, patterns exhibited during diagnostic search, and web content and anxiety during such searches.

“Through a better understanding of intentions and attention in how people find medical information online,” White concludes, “we hope to inform the design of health-search technology in search engines. Ultimately, we hope this will help people make better decisions on their own health seeking and guide engagement with medical professionals.”