Summarized by Christian Herzog
Highlights of the VEIL
In this Virtual Ethical Innovation Lecture, Kirsten Ostherr, Gladys Louise Fox Professor of English at Rice University, Houston, Texas, USA, talks about the challenges of unstructured data — a vital source of information in narrative medicine — for bringing the promises of AI in healthcare to fruition. Her presentation is structured into three parts, first outlining the issue of unstructured data for AI in healthcare, then discussing the opportunities and challenges of natural language processing (NLP) in narrative medicine, and finally tackling the subject of the social determinants of health and how AI may address or compound disparities.
Narrative as “Unstructured Data”
Kirsten Ostherr commences by outlining the phenomenon of AI hype in healthcare. It was only until after 2020 that tangible results had actually found their way into practice. This had immediately spawned anxieties about job displacement. However, before 2020, the path towards effective AI in healthcare was ridden with disappointments. IBM cancelled a leukaemia-related project in 2016 after having spent more than 62 million USD on it. But even after that, in 2022 IBM Watson sold most of its data sets, spawning further doubt on AI to be able to generate lasting economic value generation and improved health outcomes.
In essence, Strickland (2019) summarizes the issue: The promise of medical AI was built on carefully constructed lab experiments. However, up to 80% of data in electronic health records are comprised of unstructured textual input.
Natural Language Processing and Narrative Medicine: An Opportunity
By recognizing that patient stories are a vital component of the medical record, Kirsten Ostherr makes a point that these stories should be included into medical documentation rather than replaced by generic, drop-down menu items and fixed ontologies.
Indeed, major health organizations around the world have continually emphasized the need to incorporate patient perspectives in healthcare and research (Snyder et al., 2013). Natural language processing may in fact provided one way to better integrate qualitative and quantitative data in patient records. The corresponding NLP methods would benefit from the interpretation of narratives in complex intersubjective, social, and cultural contexts — these should not be bracketed out (Dohan et al., 2016). There is even calls for using NLP to integrate the analysis of clinical records (including unstructured free text) with user-generated content from social media (Martínez et al., 2016).
For instance, NLP based text mining on Yelp datasets of over 16.000 reviews of US hospitals indicated that more relevant items are covered in these social media-type reviews than in the official HCAHPS patient satisfaction survey (Merchant et al., 2016).
Kirsten Ostherr is eager to clarify that these approaches incur significant privacy and fairness implications, among others. However, a major point shed outlined consists in the opportunity to detect large-scale health trends by augmenting purely clinical data with insights from unstructured data outside the clinical realm. In fact, many larger tech companies have long understood that the lack of regulation on this data makes its acquisition relatively easy compared to clinical data.
Social Determinants of Health: AI Addressing or Compounding Racial Disparities
The social determinants of health (SDH) are the non-medical factors that influence health outcomes. The World Health Organisation defines them as „the conditions in which people are born, grow, work, live, and age, and the wider set of forces and systems shaping the conditions of daily life.“ This stands to recognize that health and illness occurs largely outside of clinical settings, where abundant meta-clinical user-generated health data can be harvested. Hence, social media feeds may potentially speak to, e.g., how individuals respond to racism, etc. The fact that this data is messy, voluminous and heterogeneous seemingly creates a demand for tools like AI.
Meta-clinical data can be harvested from a variety of sources and most consumer-facing apps & devices face no regulation, e.g., by the Food & Drug Administration. Hence, outside the clinical setting, tech companies can exploit practices of health datafication. Regulation like the GDPR offers some mild protection, but violations continue also in part because enforcement is difficult. In effect digital profiling and data mining practices are driving an exploitation of the concept of social determinants of health, as largely publicly available data can easily produce scores for medical risks.
However, social media platforms have also done significant good as online patient communities have adopted them to organize, support and educate themselves. This way, they are generating new forms of health datafication through practices of self-narration and digital connection, with real benefits.
The question then remains, how humanities can inform narrative data-mining bypassing ethical issues while maintaining or reaping new benefits.
Q&A
An extensive Q&A session revolved around alternative approaches to improve the impact of narratives in medicine. For instance, Kirsten Ostherr provided some background on the use of patient stories (Charon, n.d.). In addition, she mentioned the Open Notes Movement, which gives patients insight into the exact health records with positive effects (Blok et al., 2021). The notion that rather subtle changes may have substantial effects on doctor-patient relationships was brought up, which Kirsten Ostherr suggests should be framed from a design perspective. Further discussion points touched on medicalization and issues of hype both in AI and AI ethics. The importance of fusing perspectives on specific ethical issues in AI, such as algorithmic bias, with overarching ethical considerations in business and marketing were discussed.
Literature
Blok, A. C., Amante, D. J., Hogan, T. P., Sadasivam, R. S., Shimada, S. L., Woods, S., Nazi, K. M., & Houston, T. K. (2021). Impact of Patient Access to Online VA Notes on Healthcare Utilization and Clinician Documentation: A Retrospective Cohort Study. Journal of General Internal Medicine, 36(3), 592–599. doi.org/10.1007/s11606-020-06304-0
Charon, R. (n.d.). The sciences of narrative medicine. 3.
Dohan, D., Garrett, S. B., Rendle, K. A., Halley, M., & Abramson, C. (2016). The Importance Of Integrating Narrative Into Health Care Decision Making. Health Affairs, 35(4), 720–725. doi.org/10.1377/hlthaff.2015.1373
Martínez, P., Martínez, J. L., Segura-Bedmar, I., Moreno-Schneider, J., Luna, A., & Revert, R. (2016). Turning user generated health-related content into actionable knowledge through text analytics services. Computers in Industry, 78, 43–56. doi.org/10.1016/j.compind.2015.10.006
Merchant, R. M., Volpp, K. G., & Asch, D. A. (2016). Learning by Listening—Improving Health Care in the Era of Yelp. JAMA, 316(23), 2483. doi.org/10.1001/jama.2016.16754
Snyder, C. F., Jensen, R. E., Segal, J. B., & Wu, A. W. (2013). Patient-reported Outcomes (PROs): Putting the Patient Perspective in Patient-centered Outcomes Research. Medical Care, 51(Supplement 8Suppl 3), S73–S79. doi.org/10.1097/MLR.0b013e31829b1d84
Strickland, E. (2019). IBM Watson, heal thyself: How IBM overpromised and underdelivered on AI health care. IEEE Spectrum, 56(4), 24–31. doi.org/10.1109/MSPEC.2019.8678513