TL;DR
AI in primary care is often developed in a vacuum, leading to tools that don't fit clinical workflows. This study uses "deliberative dialogues" with Canadian patients and doctors to flip the script, revealing a strong preference for AI that automates administrative "drudge work" (like documentation and triage) over high-stakes diagnostic tools, which are currently perceived as too risky.
What Problem Does This Paper Address?
While AI is making waves in image-heavy specialties like radiology, its integration into the "front lines" of medicine—primary care—remains stagnant. Developers often focus on complex Clinical Decision Support (CDS) and diagnostic algorithms, but these tools frequently ignore the unique, relationship-based nature of family medicine.
The authors identify a significant "implementation gap" driven by three factors:
- Misaligned Priorities: A disconnect between what engineers build and what clinicians actually need to manage burnout.
- Trust and Safety: Deep-seated concerns regarding algorithmic bias, the "black box" nature of deep learning, and the potential for "deskilling" future physicians.
- The Human Element: The fear that AI might disrupt the sacred patient-provider relationship, which is the cornerstone of primary care.
Methods
The researchers employed a qualitative approach called Deliberative Dialogue. Unlike a standard focus group, this method involves providing participants with evidence-based information (via an "Informational Module") before engaging them in structured, multi-round discussions to identify shared priorities.
- Participants: 48 stakeholders (22 patients, 21 providers, 5 health system leaders) from 8 Canadian provinces.
- Process: 12 virtual dialogues held over three rounds. Round 1 focused on identifying priorities; Round 2 explored impact through use cases; Round 3 addressed barriers and consensus.
- Analysis: Dialogue transcripts were analyzed using interpretive description, a qualitative strategy used to generate practical, clinical insights.
Datasets
As this is a qualitative study on policy and priority-setting, it does not involve training a machine learning model on a specific dataset. Instead, the "data" consists of the transcribed dialogues of the 48 participants.
However, the paper discusses the types of data that primary care AI should leverage, specifically highlighting:
- Electronic Health Records (EHRs): Longitudinal data spanning diseases and socioeconomic circumstances.
- Administrative Data: Billing and scheduling information.
- Patient-Generated Data: Information from wearables (heart rate, glucose levels).
Key Insights and Ideas
- The "Administrative First" Mandate: Both patients and providers reached a striking consensus: AI should first be applied to "low-risk" tasks like automated charting, referral management, and triage. The goal is to "liberate" the doctor from the screen so they can actually look at the patient.
- High Reward vs. High Risk: While participants saw the potential of AI-driven diagnostics (e.g., predicting hospitalizations), they categorized these as "lower priority" for now. The perceived risks of algorithmic bias and lack of transparency (the "black box" problem) currently outweigh the perceived benefits.
- The Fear of "Deskilling": A recurring concern was that future generations of doctors might lose core clinical reasoning skills if they become overly dependent on AI "crutches."
- Co-Design is Non-Negotiable: The authors argue that AI developers must move beyond "building for" primary care and start "building with" stakeholders as co-design partners from the very first line of code.
Key Results
- Priority Ranking: Documentation, practice operations, and triage were identified as the highest-priority areas for AI application.
- Shared Values: Health equity, patient safety, and "person-centered" delivery were the three core values that stakeholders insisted must guide AI implementation.
- Training Needs: The study identified a critical need for "AI literacy" in medical education, focusing on algorithm critical appraisal and understanding medicolegal liability.
- Shift in Scope: One physician participant described the ideal impact of AI not as "full scope" (doing everything) but as "optimum scope" (allowing humans to do what they do best).
Limitations and Future Work
Limitations: * Selection Bias: Participants were generally highly educated and already interested in technology. * Representation Gap: The study lacked self-identifying Indigenous participants and individuals from Northern territories, whose unique resource constraints might lead to different priorities. * Language: Dialogues were conducted only in English.
Future Work: The authors propose an "agenda for the future" where AI developers work in interdisciplinary teams that include health information specialists and implementation experts. Future research should focus on creating "explainable AI" (XAI) that allows clinicians to understand why a prediction was made, thereby building the trust necessary for higher-stakes clinical applications.
Summary updated