Expanded Executive Summary

“There is surely nothing quite so useless as doing with great efficiency what should not be done at all.” — Peter Drucker

Context

In high-acuity nursing environments (e.g., ICU, medical-surgical), the accurate recording of inpatient data is not only an essential component of quality healthcare, but a physical communication tool among members of a patient’s care team. In hospitals across the U.S., despite the prevalence of charting — i.e., entering short-form patient data into EHRs (electronic health records, typically by Epic or Athena) — there is ample evidence that nurses rely on other artifacts, typically paper, to collect and recall patient data for charting purposes.

While physicians have successfully adopted speech recognition technology to record patient data through narrative-like dictation (considered “documentation”), nurses’ primary use cases (considered “charting”) remain markedly different.

Business Goal

Explore the feasibility of leveraging the client’s high-touch AI-powered technologies (DMO, DAX) for inpatient charting.

Research Goals

• Understand nursing workflows for patient data collection

• Capture and explore their specific goals & motivations

• Identify obstacles that prevent the realization of these goals & motivations

• Propose and lightly validate concepts to help them overcome these obstacles

• And, in so doing, understand their occupational culture — the waters in which they swim — by referencing definitions of culture older and newer. For instance:

  • Geertz, 1973: “A set of control mechanisms — plans, recipes, rules, instructions (what computer engineers call “programs”) — for the governing of human behavior.”

  • Jones, 2023: “The human capacity to share and learn information, knowledge, skills, stories, emotions, and innovation capabilities that empower human groups to adapt and thrive in their unique ecosystems.”

Protocol / Methodology

• A literature review of 25 journal articles about nursing workflows and the uses of voice tech in nursing to assess the current technological and attitudinal landscape.

• 3 client stakeholder interviews to align concerned parties, address information asymmetry, and pressure test research questions.

• 11 semi-structured cognitive interviews. Recruitment followed suit, which relied upon a convenience sample that had passed a screener.

  • Cognitive interviews enhance memory recall, reconstruct context, provide detailed descriptions, elicit emotions, and detect loaded language

• Using Dovetail, data were coded in several phases in a shared code book. After surfacing early patterns in an initial grounded theory phase, successive rounds of thematic analysis assigned structural and process codes to all interviews.

  • Why code in the first place? Codes are knowledge objects, and matching code types gives systematic form and flow to the data. Moreover, in-depth, de-biased analysis was ultimately achieved by having members code interviews they neither moderated nor took notes for. This process yielded 100+ discrete codes, 800+ coding instances, culminated in 40 core findings, 24 distinct insights, and helped to avoid contested vision.

• To balance outsider (client) vs. insider (nurses) perspectives (etic vs. emic), I added elements of an ethnomethodological workplace study in the form of interpretative phenomenological analysis, including epoché procedure:

  • To contextualize lived experiences: e.g., perspectives, attitudes, values, beliefs, behavioral norms

  • To understand how participants perceived, constructed, ordered, made sense of, and acted upon their experiences and social realities

    • This also relates to Umwelt, which describes the unique and subjective ways people experience their surroundings based on their sensory, motor, and cognitive abilities

  • To facilitate theoretical development, which is far from an academic luxury, as it bolsters explanatory power, informs practice, and organizes reality

    • “Experience by itself teaches nothing. Without theory, experience has no meaning. Without theory, one has no questions to ask. Hence, without theory, there is no learning.” — W. Edwards Deming

Findings → Themes → Insights

With one exception, as filtered through Figure 1, 91% of participants expressed ranges of ambivalence, reticence, and hostility. Reasons fell along several lines:

• Unlike physicians who keep narrative-like records, nurses typically capture short, discrete bursts of information, which are additionally used as memory assists (which is not surprising, as writing forges stronger neural pathways.) This population didn’t view speech recognition tech as conducive to charting and note-taking.

• For privacy reasons, participants couldn’t conceive of taking voice notes in front of loved ones, who are oftentimes in the room.

• They negatively viewed the idea that (conscious) patients should be audibly exposed to jargon being spoken into an EHR.

• While working in EHRs may still leave something to be desired — they have their own workflow issues — the prevailing attitude was: “Better the devil you know.” And while paper notes have minor shortcomings, they remain workable solutions. Moreover, they’re effortlessly passed among nurses during shift changes and don’t suffer from glitches or connectivity issues. Most fundamentally, these notes are deeply ingrained behavioral habits not easily abandoned, and this tech was largely considered yet another digital tool to interfere with nurse-patient interaction.

  • This major insight relates to function allocation as well as the “ironies of automation,” which was presciently described in 1983 (and regularly since).

• Interestingly, a minor finding worth reporting was a lack of concern surrounding the quality of speech recognition.

So What? Impact and Business Outcomes

The primary conclusion is that this technology represented solutions in search of problems. Nuance subsequently abandoned the line of inquiry after triangulating results with related research that held similar findings. With independent studies validating each other, this became a story of opportunity cost, given how this inquiry saved the company further resources by avoiding investment (money & time) that targeted a population that demonstrated signs of being unreceptive.

Put another way, and not without irony, it would have been iatrogenic (i.e., care-induced injury) for a medical technology company to have disregarded these consistent findings by pursuing such a “solution.”

“Almost all men die from their drugs and not from their diseases.” — Molière, 1622-1673, The Imaginary Invalid, Act III

Next Steps / Opportunities / Recommendations

Despite the above results, speech recognition technology is more than just voice dictation. Final recommendations regarding its use in this context:

• Investigate the potential intersection of speech recognition technology with communication devices already used in these settings, if not for charting, then at least for note taking.

• Investigate potential correlations surfaced through the participant spectrum map (Figure 2). Nurses who work at hospitals that require structured, templated note-taking also reported that charting has a “very high impact” on work-life balance (due to EHR workflow issues).

• Identify clinics that don’t allow paper into the patient’s room due to COVID-19. If the practice is enforced in more inpatient settings than what was found with this convenience sample, there may be a perceived utility for voice-enabled note taking.

• Continue exploring ways to replace paper as a stopgap solution for quickly adding highly structured information to an EHR system. Utilize a participatory design method like a charrette with nurse participants.

• Continue the line of inquiry with the next generation of nurses because, while still skeptical, the youngest practitioners were at least more entertaining of the idea.

Limitations

• Participant observation and interviews in situ would have been ideal, which likely would’ve been possible in a pre-COVID world. Regrettably, it wasn’t an option at the time of this study.

• Recruitment fell a bit short of the 12 minimum participants desired. (Why 12? See here and here, if interested.)

  • Patient technicians, who also play a role in high-acuity patient care, would have been welcome. But recruitment limitations unfortunately did not allow for this.

• Coding and analysis took longer than expected, which led to a deviation from the protocol’s final step: a charrette with nurses. However, there was time to involve client stakeholders in a blue-sky ideation activity, which yielded welcome out-of-the-box solutions.

Retrospective

• What did I learn?

  • An abundance of methodological skills applied to a critical subject area.

  • There are signs that nursing attitudes will eventually, albeit slowly, evolve. It’s currently a slow-moving train but I’m curious to see how attitudes shift as NLP and AI catch fire in new ways each year.

• What would I have done differently?

  • Been less granular with building the codebook, to save time.

  • Maintained a better sleep schedule.

• What went well? What will I keep doing?

  • The client was deeply engaged and supportive at every turn.

  • Participants were gregarious, needing little prompting to be openly expressive.

  • I kept my cool and nose to the grindstone. A demanding schedule tried collective patience, but it was sink or swim and the focus was to stay afloat.

  • To that end, team cohesion, candor, and engagement were paramount to success. This unity provided psychological safety, productivity, and interpersonal satisfaction. All primary objectives were meaningfully met.