The views of health guideline developers on the use of automation in health evidence synthesis

Anneliese Arno, Julian Elliott, Byron Wallace, Tari Turner, James Oscar ThomasResearch Square 2020 · DOI

TL;DR

While machine learning could theoretically solve the "information overload" in medical research, health guideline developers are hesitant to adopt it. This study reveals that the primary barrier isn't technical complexity, but a perceived clash with professional values like human judgment and rigor. For automation to succeed in this field, it must be designed as a transparent tool for human redirection, not a "black box" replacement for expert insight.

What Problem Does This Paper Address?

The pace of medical research is staggering, with nearly 4,000 health-related articles published every single day. This "deluge of data" makes it nearly impossible for researchers to keep systematic reviews and clinical guidelines up to date using traditional manual methods.

While automation technologies—such as machine learning (ML) and natural language processing (NLP)—exist to speed up evidence synthesis, their adoption has been remarkably slow. The authors set out to understand why. By interviewing the "gatekeepers" of medical evidence—the people who develop the guidelines that doctors actually follow—the researchers aimed to identify the specific barriers and facilitators to integrating AI into the evidence pipeline.

Methods

The researchers conducted a qualitative study using semi-structured interviews with 18 guideline development experts from major international organizations (including NICE in the UK, NHMRC in Australia, and the WHO).

To analyze the data, they applied Rogers’ Diffusion of Innovations framework. This classic sociological model evaluates how new ideas spread based on five dimensions:

  1. Relative Advantage: Is it better than what we have now?
  2. Compatibility: Does it align with our values and needs?
  3. Complexity: Is it difficult to understand or use?
  4. Trialability: Can we experiment with it first?
  5. Observability: Are the results visible and verifiable?

The analysis used a "mixed deductive and grounded approach," meaning they first categorized responses into these five buckets and then looked for organic sub-themes within them.

Datasets

As this is a qualitative social science study, it does not use traditional computational datasets. Instead, the "dataset" consists of 18 verbatim interview transcripts from experts with varying levels of experience (ranging from less than 5 to over 20 years) in evidence synthesis across government, academic, and private sectors.

Key Insights and Ideas

  • Compatibility is the Ultimate Hurdle: The most significant finding was that "Compatibility" dominated the conversation (accounting for roughly 45% of the thematic coverage). Developers aren't worried about the software being "hard to use"; they are worried that it doesn't align with the "rigorous" and "human-centric" culture of medicine.
  • The Fear of the Black Box: Participants expressed a deep need for Observability. They don't just want an accurate result; they want to see the "inner workings" of the algorithm. Transparency isn't just a feature; it’s a requirement for accountability.
  • Redirection vs. Replacement: There is a strong consensus that AI should redirect human effort rather than replace it. Developers are happy to let a machine handle the "drudge work" of screening thousands of papers, provided it frees up humans to perform the nuanced, high-level "value decisions" that machines currently cannot mimic.
  • Accuracy is Non-Negotiable: Unlike other fields where a "speed-accuracy trade-off" might be acceptable, guideline developers showed zero openness to sacrificing accuracy for time. If a machine is 95% accurate but a human is 99%, the machine is viewed as a risk to population health.

Key Results

  • The "Value" Gap: The study found that developers prioritize "alignment with values" (e.g., the ability to double-check work) over "alignment with infrastructure."
  • The Adoption Curve: Based on the interviews, the field of health guideline development is currently in the "Innovator" stage of the adoption curve. Most participants felt they lacked robust knowledge of what automation can actually do.
  • Quantitative Thematic Distribution:
    • Compatibility: ~45% coverage.
  • Relative Advantage & Observability: ~25% each.
  • Complexity & Trialability: ~5% each (indicating these are minor concerns compared to trust and values).

Limitations and Future Work

The authors acknowledge a few key limitations:

  • Sample Bias: The sample was relatively small (18 people) and skewed toward professionals in Australia, the UK, and the US. It also had a high percentage of female participants (72%), though the authors noted that views did not seem to vary by gender.
  • Stakeholder Scope: This study focused strictly on guideline developers. Future research needs to include the users of the software (systematic reviewers) and the beneficiaries of the guidelines (patients and policy makers).

Future Direction: The paper suggests that proponents of AI in medicine should stop selling "speed" and start selling "transparency." Future work should focus on creating "human-in-the-loop" systems that emphasize how automation supports, rather than replaces, expert judgment.

Summary updated

Abstract

<title>Abstract</title> Background: The increasingly rapid rate of evidence publication has made it difficult for evidence synthesis – systematic reviews and health guidelines -- to be continually kept up to date. One proposed solution for this is the use of automation in health evidence synthesis. Guideline developers are key gatekeepers in the acceptance and use of evidence, and therefore their opinions on the potential use of automation are crucial. Methods: The objective of this study was to analyze the attitudes of guideline developers towards the use of automation in health evidence synthesis. The Diffusion of Innovations framework was chosen as an initial analytical framework because it encapsulates some of the core issues which are thought to affect the adoption of new innovations in practice. This well-established theory posits five dimensions which affect the adoption of novel technologies: <italic>Relative Advantage</italic>,<italic> Compatibility</italic>,<italic> Complexity</italic>, <italic>Trialability</italic>,<italic> </italic>and<italic> Observability</italic>.Eighteen interviews were conducted with individuals who were currently working, or had previously worked, in guideline development. After transcription, a multiphase mixed deductive and grounded approach was used to analyze the data. First, transcripts were coded with a deductive approach using Rogers’ Diffusion of Innovation as the top-level themes. Second, sub-themes within the framework were identified using a grounded approach.Results: Participants were consistently most concerned with the extent to which an innovation is in line with current values and practices (i.e. <italic>Compatibility</italic> in the Diffusion of Innovations framework). Participants were also concerned with <italic>Relative Advantage</italic> and <italic>Observability</italic>, which were discussed in approximately equal amounts. For the latter, participants expressed a desire for transparency in methodology of automation software. Participants were noticeably less interested in <italic>Complexity</italic> and <italic>Trialability</italic>, which were discussed infrequently. These results were reasonably consistent across all participants. Conclusions: If machine learning and other automation technologies are to be used more widely and to their full potential in systematic reviews and guideline development, it is crucial to ensure new technologies are in line with current values and practice. It will also be important to maximize the transparency of the methods of these technologies to address the concerns of guideline developers.
BibTeX
@inproceedings{arno2020views,
  author = {Anneliese Arno and Julian Elliott and Byron Wallace and Tari Turner and James Oscar Thomas},
  title = {{The views of health guideline developers on the use of automation in health evidence synthesis}},
  booktitle = {Research Square},
  year = {2020},
  doi = {10.21203/rs.3.rs-23742/v2},
}
Figure