Using deep learning to improve medication safety: the untapped potential of social media

In order to deliver patient-centered healthcare, knowledge of the patient experience is required. Unfortunately, the majority of data-driven solutions in healthcare to date have not included patient-reported information. Social media is a popular and growing platform on which health-related information is routinely shared (e.g. Twitter, PatientsLikeMe, WebMD). Yet, only a small percentage of such information is being used to improve patient care.

Moreover, big-data techniques are rapidly advancing. While other industries have successfully integrated big data-derived functions from text mining, healthcare has only recently begun doing so. Text mining has under-delivered on providing meaningful action in healthcare settings largely due to noisy patient-generated text including spelling errors, imprecise descriptions, and misused medical terms. As a result of the advantage of representation learning and deep architecture, recent developments in deep learning have shown tremendous promise that technological barriers to text mining can be overcome. ¹

While promising for use in acute health situations, social media contains untapped insights on the experiences of millions of patients with chronic diseases. ² In the USA, about half of all adults, that is, 117 million people, have at least one chronic condition. ³ Importantly, more than 30% of patients are willing to share information about their health with other patients on social-media sites. ⁴ The most common treatment for chronic disease is medication therapy, and medication safety is, in turn, a top priority for public health. Two of the most pressing medication safety issues are medication nonadherence and adverse drug events (ADEs), phenomena that require patient-reported information. We believe that big health data analytics with deep-learning techniques can be fruitfully leveraged to uncover new and actionable insights about medication safety.

Medication adherence

Medication nonadherence is widely recognized as a major public-health problem, affecting patients of all ages and health conditions. More than 50% of adults do not adhere to their chronic medications, accounting for US$290 billion in preventable costs annually. ⁵ One of the largest barriers to understanding and improving medication adherence is the challenge of measuring a complex health behavior. Over a decade ago, Osterberg and Blaschke articulated various methods of measuring adherence. ⁶ However, social media was not listed as a method because it was not ubiquitous at the time. Today, the vast majority of Americans use social media every day. Moreover, patient self-reports of medication adherence have been consistently critiqued for risk of social desirability and recall biases. However, these biases are much less of an issue on social-media platforms. Due to the variety and uncertainty of patient medication-taking behavior, research into medication nonadherence can greatly benefit from new measurement methods and patient-generated data sources. With the growth and accessibility of health-related social-media data, we propose that a big data-based approach can be effectively leveraged to complement traditional methods (e.g. pharmacy claims, pill count) to advance a variety of medication-adherence research streams.

Social-media analytics with deep-learning techniques could advance the field of medication adherence in the following ways: (a) tailoring solutions to medication nonadherence for individual patients; (b) improving our understanding of why patients do not adhere to medications (including prescription and over-the-counter); (c) enhancing our ability to predict who is at risk for being nonadherent; (d) broadening effective interventions for improving medication adherence. First, prior studies have rarely considered the characteristics of each nonadherent patient and features of individual drugs. For this reason, the US National Institutes of Health are promoting a new medical model, that is, precision medicine, that takes into account individual variability in response to treatments. Deep-learning methods could enable such precision measurement for individual patients. Second, healthcare providers are unable to identify accurately nonadherence, so further study is needed to analyze patient-generated text for adherence barriers in a more ecologically valid setting (i.e. online) than the clinical encounter. Our recent study has successfully developed a deep-learning framework to examine social-media text and identified individualized reasons for medication nonadherence. ⁷ Third, by gaining insights into the patient-medication experience, predictive models of nonadherence can be built more accurately. A similar approach was implemented by HealthTap, which introduced Dr A.I. to help triage patients based on symptoms. Fourth, patient-generated text can provide novel insights into patient-identified strategies for improving medication adherence.

ADEs

ADEs are common and costly in the USA, affecting 20% of the population and leading to an estimated US$30.1 billion in preventable costs annually. ⁸ Although postmarketing monitoring systems, such as the US Food and Drug Administration’s (FDA) MedWatch, have been established to collect reports of ADEs, most patients are not aware of such systems. Consequently, a significant number of ADEs have never been reported. There is a great need to explore cost-efficient reporting channels for ADEs. ⁹

The pharmaceutical industry has started to monitor health topics on various social-media platforms, such as Epidemico and WEB-RADR. They have developed online platforms to collect and analyze disease outbreaks, food and drug safety, and patient experiences. Our recent research improved this line of work by developing a deep-learning framework to detect ADEs of e-cigarettes. We examined 6 million social-media postings and identified 1591 unique e-cigarette-related ADEs, including many not noted by prior studies, such as allergy, eye-twitch, fatigue, and asthma. ¹⁰ Our deep-learning technique improves upon conventional text-mining methods by addressing the informal and nontechnical consumer vocabulary issues. In the test setting, our approach was able to extract 91.8% of all ADEs with a precision of 94.1%.

With deep-learning techniques, pharmacovigilance from social media could be advanced in the following ways: (a) by improving the accuracy of detecting ADEs in large-scale, patient-generated texts; (b) detecting ADEs while accounting for the heterogeneity in patients’ demographics; (c) exploring the adverse outcomes of various drug interactions, such as drug–drug interactions, drug–supplement interactions, and drug–food interactions.

Challenges and opportunities

To realize fully the potential of a big data-based approach to medication safety, several challenges need to be addressed. First, deep-learning techniques require large-scale examples containing the ground truth (i.e. previously known output). For instance, a predictive model for medication nonadherence needs thousands of text records with known nonadherence. Preparing the ground-truth data requires intensive human annotation. To maximize the impact of such research, standardized datasets and research collaboration will be required. Second, although patient-generated text contains valuable insights about patients’ medication experiences, the content may contain emotional and biased statements. Uncovering the true medical circumstances can be challenging. Thus, this research requires a multidisciplinary framework with active participation from clinicians, researchers, patients, and text-mining experts. Third, as knowledge discovered via big-data methods tends to focus on correlations (as opposed to causation), the findings need to be provided in the appropriate context. In this sense, big-data approaches serve to complement, rather than replace, the current methods and channels for collecting medication safety information. Fourth, a smaller portion of older adults than younger adults uses social media, and people in underdeveloped countries may use social media less frequently. To address this issue, large representative and multi-language datasets need to be constructed and analyzed.

We expect fast growth of deep-learning methods in the coming years. A major opportunity exists to utilize the vast amounts of health-related data on social media to better understand and ultimately improve medication safety. Patients are talking. Are we listening?