Clinicians play an essential role in the development of digital health applications by bringing necessary medical domain expertise to the table. Thanks to the current digital health revolution, interest in healthcare AI is at an all-time high amongst clinicians. Interest, however, is only half the battle, and several significant hurdles remain with respect to development and deployment toolsets for digital health systems that maximizes clinical investigator engagement and fully leverages supplied clinical expertise.
Usage of unstructured data is indisputably critical for successful implementations of digital health systems, particularly in the context of a learning health system. Over a decades-long course of applying NLP in support of development and deployment of such a system, we at the Mayo Clinic have gone through a large variety of toolsets and approaches, starting from the widely known Clinical Text Analysis and Knowledge Extraction System (cTAKES) to our most recent NLP as a Service (NLPaaS) platform presented in this article.
Many of these changes in NLP approaches are adaptations resulting from challenges encountered from various projects related to our goal of a learning healthcare system: from bottlenecks encountered during the research and development phase, to issues pertaining to production deployment, to barriers towards widespread adoption and use of developed algorithms.
In our article, we highlighted several of these issues, particularly in the context of the development of cross-institutionally portable and deployable healthcare AI, and formulated these issues as three general desiderata for the development of an NLP platform. Also contained therein is an example implementation of the NLPaaS platform, which realized these desiderata as user-friendly, high-throughput, and collaborative NLP self-service platform for clinician collaborators.
Of particular note, we envision a collaborative crowd-sourcing approach to deal with the semantic portability gap. Given the wide variety of semantic perspectives held by differing clinicians from differing institutions, we believe that, to create a robust clinical AI, it is critical that all these perspectives be considered given decision making. Given sufficient example perspectives (via crowd-sourcing), we theorize that traditional statistical approaches can learn precisely how to weight these different perspectives given any input data-set.
Will such an approach work? Time will tell: we certainly believe so, but it is impossible to test without a critical mass of semantic perspectives - thus necessitating the development of our proposed NLP platform. Such an experiment on cross-institutional portability is, of course, not fully doable by only a single institution: we gladly welcome you (and the perspectives you bring) to join us on our journey.
We are excited to share the lessons we have learned and cordially hope they serve a constructive purpose. We invite the community to participate in a dialogue by sharing your valuable feedback alongside your own experiences on the topic.