Ido Hadari, Treato

Fig. 1. The social media inputs needed to get to useful information for

						determining patient outcomes. Credit: Treato

		It is widely recognized that social media is a powerful forum for patients and providers to share views about healthcare. The FDA even recently issued a first set of proposed guidelines that acknowledges social media’s growing critical role as a source of open communication and patient interaction. Today one out of three US adults go online for information, support and advice about their health; more and more patients are coming to their physicians with opinions formed about what therapies they might need, what drugs address those needs, and how well a therapy is progressing—the entire patient journey.

		These online conversations have important implications for biopharma. Through the crop of available “social listening” tools, pharma can “listen in” but the latest technology is helping them go even farther to new depths of understanding about the empowered online patient. Such technology can now turn the massive amounts of disparate online patient conversations into one structured “patient voice” that can significantly impact healthcare’s product and business strategy—from brand lifecycle management to market intelligence and education to R&D.

		The key is knowing where and how to look. And getting past a few prevailing notions about what social media is all about when it comes to the pharmaceutical industry. This article attempts to provide some useful guidance on charting the often murky social media waters to reach the empowered patient online. Here are five key things to consider as you make your way through that patient kingdom:

		5) Be strategic: it’s an enterprise play

		Getting the most out of social networks means going far beyond Facebook or Twitter. There are literally thousands of different places for patients to interact, comment and share on the social health web. What’s known as the “long tail” of the Internet consists of scores of interesting online destinations, focusing on specific therapeutic categories, patients with certain shared backgrounds, areas of interest or geographical location.

		Tackling this breadth and scale is a serious technical challenge, yet is critical to rendering the patient voice comprehensively. As an example, at Treato, where we gather, aggregate and analyze patient-generated content on the web, we sift through and process around one million healthcare related posts daily—from roughly 3,000 different health-related sites. These figures are on a continuous growth trajectory.

		Making sense of such large scale, dynamic patient content is complex. A crucial first step requires identifying all the user generated content (UGC) that includes health-specific content, and then eliminating the “noise.” As FDA guidelines now suggest, this independently generated patient content is at the heart of what social media in healthcare is all about. But to make sense of it, there is a need, first, to filter commercial content and spam, which can account for as much as 80% of online content.

		Another factor to consider is hearsay, that is, posts that talk about what someone heard or read about a treatment or disease, as opposed to something they experienced firsthand. For example “I started feeling depressed after taking the drug” is very different from “I heard that this drug can cause depression.” Therefore, including hearsay in any social media analysis can significantly compromise its reliability and create bias throughout the decisionmaking process.

		The foundation of this sophisticated filtering is natural language processing (NLP), a method of analyzing text or speech that derives meaning and structure from UGC (see box). Once the content has been collected and filtered, only the “pure” patient- and caregiver-written posts about personal health and treatment experiences should remain.

Fig. 2. Treato’s experience in accessing patient conversations and websites to date

		Finding key messages and insights in UGC can be like looking for a needle in a haystack. The content patients and caregivers post online—like most social media updates&mdash;is typically written like a story, in a format that is technically known as “unstructured.”

		“I was diagnosed last year with MS after some vision problems and numbness. My dr. started me on Avonex, but I developed an allergy to interferon and moved to Copaxone.”

		A human being can read this patient-written post and string together a sequence of events, map the mentioned drugs to a condition, identify a side effect followed by the switching behavior. But when you are trying to analyze and distill meaning from hundreds of thousands of MS-related posts, it is humanly impossible to pick apart each unstructured post, or to connect the disparate dots across such a huge array of data. Automated methods must be implemented to convert the vast set of unstructured information into a structured, usable set.

		The right approach doesn’t just automate this process but can translate this unstructured content into meaning based on well-defined reference information or fields. In the case of healthcare, such fields can consist of “switching behavior,” “side effects,” “initially diagnosed,” and so forth. In this sense the unstructured content becomes structured.

		However, to achieve that, extensive medical expertise must be employed, and since this task is conducted by machines that expertise should be an integral part of any algorithms deployed to decipher a patient’s post. In this case, health-specific ontologies will inject the specific domain knowledge into the mix: drugs, treatments, indications, symptoms—and the relations between them.

		Patients and caregivers also talk (and write) in a vernacular that can be quite different from the language HCPs use when talking about medication and conditions. For example, patients may just as easily use the term “heart attack” when expressing an emotion of shock, and not mean “myocardial infarction.” Ambiguity arises frequently within healthcare-related content. Abbreviations like “BP” are a representative example, referring to Blood Pressure in one context, Bi-Polar in the context of a mental health forum, or British Petroleum in a non-health forum. At the same time, patients may use different terms to express the same thing—a patient complaining about “leg pain” may be experiencing the same thing as another patient describing their “calves hurting.”

		Well-organized NLP techniques, algorithms and vocabularies create structure and meaning out of this content.

		When it comes to pharma and the vast amount of UGC on the social health web, merely tracking brand mentions is not enough. Through social health analytics, it is possible to drive strategy throughout the brand lifecycle and product development process by answering such key questions as:

			Where is my brand losing or gaining market share, and more importantly, why?

			What are the reasons driving drug starts, switches and stops?

			How can I improve adherence and compliance?

			What are my patients’ biggest concerns, fears and hopes?

			How are physicians impacting my patients’ decisions? What gaps or misperceptions are driving undesired behavior?

		With the right tools, pharmaceutical companies can apply analytics to social media content to gain insights into various patient segments—before, during and after a prescription is written. This can include insights about key characteristics, emotional triggers and what specific language patients are using when they talk about their conditions or medical concerns. An additional benefit is a clearer picture of the digital therapeutic landscape, to better understand where specific patient segments are sharing their experiences online, in order to better understand how to reach them.

		It goes without saying that patients experience different feelings toward a treatment method or medication depending on their different emotional triggers. Social media can unveil patient attitudes toward disease progression as they discuss their attitude change over time, for example as their condition gets better or worse, or as they move from being newly diagnosed (and feeling “hopeful,” or “uncertain”) to being long-term patients adjusting to or considering various medications (and feeling “anxious,” “frustrated,” or “concerned about cost”).

		Traditional sentiment analysis does not differentiate such critical nuances. For this, technology needs to go beyond simple trigger-word analyses to extract deeper actionable insights. Such technology can include special algorithms working in tandem with medical ontologies and knowledge databases specifically designed to enable the deeper level of understanding required in healthcare.

Articles

How new social health analytics tools are supercharging pharma's quest for patient insight