Why Sentences Matter: Mining Clinical Narrative to Learn ‘Why’ and ‘How’ Things Happen

When I hear free text clinical notes referred to as ‘unstructured data’ the inherent presumption is that is it “not” something it otherwise should be, as if “structure’ is and should be the standard for all data.

The need for structure is not a problem of data, but the largely artificial limitations in our current capacity to analyze information. There are no doubts that efforts to extract discrete data elements from free text through techniques such as natural language processing (NLP) will add to the quantity of accessible data around any one patient or population. However, we diminish the humanity of our endeavors to provide care when we force clinical documentation into pick-lists and forms, and miss much larger and more powerful opportunities to understand health and healing when we try to reduce free text into something other – and much smaller -- than what it is.

Free text is content, and sometimes content can tell us much more than discrete data. In 2008, blogger and content management expert Seth Gottlieb wrote a few things about the differences between content and data on the web; his points are also quite relevant when considering the value of clinical notes in the form of dictated and transcribed discharge summaries, radiology and pathology reports, surgical and operative notes and free-text fields in clinical information systems:

  1. Content has a voice: it is written to communicate ideas, make a point, convince; it is personal
  2. Content has ownership: someone created the note, from their perspective of authorship as defined by their levels of authority and responsibility
  3. Content is intended for a human audience, for human senses to process
  4. Content has context; even the most objective content contains lexical, syntactic and semantic clues about where the reader should focus their attention, what was important and what was not.

While content may contain clues about “what” happened – drugs prescribed and discontinued, diagnoses made, symptoms observed, procedures performed – it is the narrative structure of content that allows us to understand “why and how” things happened to patients and the experience, processes and decisions made by participants in a patients care. It also can give us a sense of the person being cared for, their relationships with others and how they respond to the process of care. Given the transformational challenges facing healthcare professionals and systems of care and practice, it is now more than ever that we need to make an extra effort to make sense of the world we work within and see to understand “why and how” things happen to and around patients.

After more than a decade working around some of the best academic and commercial minds thinking about ‘unstructured medical data”, we have a good perspective on the strengths and weaknesses of current approaches to text analytics in healthcare.

Combine this with a deep understanding of the limits of structured data analytics to address the critical issues facing healthcare today, and I submit we need a Moonshot mindset around the concept of Healthcare Narrative IntelligenceNarrative Intelligence is an interdisciplinary construct that originated in the MIT Media Group and subsequently evolved by Michael Mateas, Phoebe Sengers and others. The idea behind a Healthcare Narrative Intelligence Initiative would be to develop completely new and powerful methods to make sense of our healthcare world and inform our wisdom about clinical processes and care delivery operations by analyzing patient-centered clinical notes and reports and encouraging a narrative component to clinical documentation. Healthcare Narrative Intelligence would allow us to make sense of what we see provide a deeper understanding of patient care and operations that could include but not be limited to:

  • Action Selection and Decision Attribution: What was the path to a specific decision and what factors affected the final decision that was made; Do these factors have any effect on outcomes?
  • Collaboration/Relational Analysis: Who was involved in these decisions, what was their relationship, what were the relationship dynamics? Would a change in relational dynamics have an effect on goal achievement or system performance?
  • Goal Setting: How are goals and associated goal metrics determined?
  • Semantic and Sentiment Analysis: What words were used to describe the patient and their care; Did subjective elements such as meaning and emotion effect process and outcomes?

With these as the foundation of Narrative Intelligence, methodologies such as predictive modeling, clustering and pattern recognition, and comparative analysis could contribute to an entirely new approach to developing performance benchmarks.

Analytics driven by the Healthcare Narrative Intelligence framework would not require the presence of protected health information, as the insights and subsequent derived wisdom would be focused at the level of systems, roles and interactions.  Health systems, hospitals and academic medical center should come together to test these models and support the development and evolution of this next generation of analytics.

© Steven Merahn, MD 2015