Governing PatientsLikeMe: Information Production and Research Through an Open, Distributed, and Data-Based Social Media Network

Abstract

Many organizations develop social media networks with the aim of engaging a wide range of social groups in the production of information that fuels their processes. This effort appears to crucially depend on complex data structures that allow the organization to connect and collect data from a myriad of local contexts and actors. One such organization, PatientsLikeMe, is developing a platform with the aim of connecting patients with one another while collecting self-reported medical data, which it uses for scientific and commercial medical research. Here the question of how technology and the underlying data structures shape the kind of information and medical evidence that can be produced through social media-based arrangements comes powerfully to the fore. In this observational case study, I introduce the concepts of information cultivation and social denomination to explicate how the development of such a data collection architecture requires a continuous exercise of balancing between the conflicting demands of patient engagement, necessary for collecting data in scale, and data semantic context, necessary for effective capture of health phenomena in informative and specific data. The study extends the understanding of the context-embeddedness of information phenomena and discusses some of the social consequences of social media models for knowledge making.

Keywords:

• aggregation
• data-based
• distributed science
• engagement
• information cultivation
• PatientsLikeMe
• self-reporting
• semantic context, social denomination
• social media

ACKNOWLEDGEMENTS

I am very thankful to the editor, the reviewers and the special issue guest editors Hamid Ekbia, Jannis Kallinikos, and Bonnie Nardi, who patiently encouraged me to take the argument further.

This research was developed as part of my PhD project, and it would have not been published without the support of numerous friends and colleagues at LSE, KCL, and the wider academic community. I am especially indebted to Jannis Kallinikos, Barbara Prainsack, Michael Barzelay, Nikolas Rose, Carsten Sorensen, Ola Henfridsson, and Sue Newell for their support and feedback. Early or later drafts have also enjoyed comments from Jochen Runde, Aleksi Aaltonen, Des Fitzgerald, Ilana Löwy, Cristina Alaimo, Paul Leonardi, Attila Marton, Silvia Camporesi, Tara Mahfoud, and the ICIS 2013 DC friends Mohammad Moeini, Koteswara Ivaturi, Brian Dunn, Ina Sebastian, Inchan Kim, John Qi Dong, and Liliana Lopez. Thanks to all of you.

Notes

1. More information can be found at www.patientslikeme.com/about

2. Estimates suggest that rare diseases affect 300 million people globally, yet no FDA-approved drugs exist for 95% of rare diseases (RARE 2014).

3. Symptom severities are captured along a NMMS (none, mild, moderate, severe) scale.

4. See Chapter 5, “On tuberculosis and trajectories,” in Bowker and Star (1999), for an stimulating discussion on the relationship between health measurements and biography.

5. As of September 2014.

6. However, even engaged patients can omit very important information because of self-reporting biases.

7. This is a form of the popular statement, “the absence of evidence is not evidence of absence” (in this formulation, attributed to the astronomer Carl Sagan; see Wikipedia 2014).

8. While the system needs to be flexible, to support different life routines and goals, on particular occasions it constrains access to specific areas of the tracking tools. For example, when a patient does not update her symptom severity scores for more than a predetermined number of days, the system will not allow her to review her symptoms data without first inputting updated symptom severity data. She will also not be able to track a new symptom before providing a new symptom data update. In this way, the system tries to force data inputs when a patient's data inputting falls below a specific threshold, thus obtaining compliance through constraint.

9. The condition categories, driving different condition history questionnaires, are infections, chronic diseases, pregnancy-related, mental health, events and injuries, and life-changing surgery.

10. This, however, is possible for only a small number of conditions. Establishing what the standard set of tools should be for a specific condition requires expensive, in-depth research. Therefore, this tends to be accomplished mainly in association with condition-specific, funded research projects.

11. For example, free form, pill, vial, etc.

12. Obviously, there are simpler conditions where it would not make sense to split the world in two. For example, it would be detrimental to divide patients into those with a “broken right leg” and those with a “broken left leg”; aggregated data provides sufficient power in this case. The same is true, but for different reasons, with generic conditions of which patients rarely get to know the type (think “flu”).

13. The five generic symptoms are anxious mood, depressed mood, fatigue, insomnia, and pain.

14. This feature, however, is limited to the minority of conditions about which the staff has had the opportunity—usually in the context of funded commercial research projects—to carry out the research required to infer the symptoms most characteristic of a patient's experience of the condition.

15. This is also possible for other medical entities such as conditions and treatments.

16. The ontological status of certain medical entities is often disputed, e.g., in the case of syndromes. Sometimes the boundary between symptom and condition is blurred and shifting. Simpler cases can be dealt with more straightforwardly, for instance when the patient has entered an entity that is clearly not a symptom, e.g., a drug.

17. For instance, “toothache cognitive impairment” is a string that can be split into two symptoms, “toothache” and “cognitive impairment,” which can then be added to the database.

18. Briefly, the difference between signs and symptoms lies mainly in who is able to observe the phenomenon in question. Scheuermann and colleagues define a sign as a “bodily feature of a patient that is observed in a physical examination and is deemed by the clinician to be of clinical significance” (Scheuermann et al. 2009, 119). For instance, a lump can be a sign: Both the clinician and the patient can easily observe it. A symptom is instead defined as “a bodily feature of a patient that is observed by the patient and is hypothesized by the patient to be a realization of a disease” (Scheuermann et al. 2009, 119). For instance, the clinician does not directly observe a symptom such as a headache. Only the patient has access to the phenomenon.

19. Importantly, patients were actually recording comorbid conditions in 25% of these cases. A cause of this was that the system could associate only one condition with each patient profile. As many chronic patients live with comorbidities, they were working around this system limitation by storing comorbidities as symptoms. When, in 2011, the system was developed to allow patients to add multiple conditions to their profile, it became better able to correctly guide this kind of data inflow. The development of a considerably more complex system, in which a patient could associate to her profile any possible combination of conditions, successfully controlled this instance of data collection creep.

20. Whereby the generic “arthritis” form was disabled, requiring patients to choose a subtype. This episode saw the organization moving the boundaries defining arthritis conditions, and consequently reshaping the patient groups and sociality created through aggregation.