Towards clinical implementation of artificial intelligence in cancer care: Concept mapping analysis of provincial workshop findings

This study utilizes concept mapping analysis of a 2024 British Columbia Cancer Summit workshop to identify stakeholder priorities for AI in cancer care, revealing a strong consensus on the high importance and feasibility of clinical benefits and efficiency gains over implementation challenges, thereby providing a structured framework to guide responsible and effective AI integration.

The Gist

Imagine you are the captain of a massive, complex ship (the healthcare system) sailing through the foggy waters of cancer care. You have a new, powerful engine on board called Artificial Intelligence (AI). Everyone agrees this engine could make the ship faster, safer, and more efficient. But before you fire it up, you need to know: Is it safe? Will it actually help? And how do we install it without breaking the ship?

This paper is essentially a group brainstorming session with the ship's crew (doctors, nurses, patients, tech experts, and administrators) to figure out exactly how to use this new engine.

Here is the story of their journey, explained simply:

1. The Big Meeting (The Workshop)

The researchers gathered 48 people from British Columbia, Canada, at a cancer summit. They didn't just sit and listen to lectures; they had a hands-on workshop. They asked three simple questions:

• What are you worried about regarding AI?
• What good things do you think AI will do?
• What should we prioritize first?

The group shouted out ideas, and the researchers wrote down 265 different thoughts. That's a lot of noise! To make sense of it, they needed a way to organize the chaos.

2. The Sorting Game (Concept Mapping)

Imagine you have a giant pile of 100 different puzzle pieces (the top ideas from the 265 thoughts). You ask a team of 13 experts to sort these pieces into piles based on how they think they fit together.

• The Result: The experts naturally sorted the pieces into two big piles (Clusters).
  • Pile A (The "Guardrails"): This pile contained all the worries and rules. Things like "What if the AI lies?", "Who is responsible if it makes a mistake?", and "How do we keep patient data private?"
  • Pile B (The "Superpowers"): This pile contained all the exciting benefits. Things like "It will save time on paperwork," "It will help spot cancer earlier," and "It will let doctors spend more time talking to patients."

3. The Scorecard (What People Liked Best)

The researchers asked the experts to rate every idea on two scales:

1. How important is this? (1 to 5 stars)
2. How easy is it to do right now? (1 to 5 stars)

The Big Discovery:
Everyone agreed that the "Superpowers" (Pile B) were both more important and easier to do than the "Guardrails" (Pile A).

• The "Easy Wins": Doctors and staff are ready for AI tools that help with daily tasks, like summarizing patient notes or organizing data. These feel like adding a new, helpful tool to a toolbox.
• The "Hard Slog": The big rules, laws, and ethical safety nets (like making sure the AI isn't biased or that data is secure) are seen as very important, but also very hard to fix right now. They require changing the whole system, not just adding a tool.

4. The "Go-Zone" Map (Where to Start)

The researchers drew a map with four squares to help leaders decide what to do next:

• Top Right (High Importance, High Feasibility): DO THIS NOW! These are the "quick wins." Examples: Using AI to transcribe doctor-patient conversations or to organize patient files.
• Top Left (High Importance, Low Feasibility): PLAN FOR THE FUTURE. These are the big, hard problems. Examples: Creating new laws for AI liability or ensuring data privacy across the whole country. These need long-term investment.
• Bottom Right (Low Importance, High Feasibility): NICE TO HAVE. Easy to do, but not a huge priority.
• Bottom Left (Low Importance, Low Feasibility): IGNORE FOR NOW.

5. The Takeaway: A Two-Step Dance

The main lesson of this paper is that we shouldn't try to do everything at once. It's like building a house:

1. First, build the rooms you can live in today. (Use AI to help with paperwork and scheduling). This gets people excited and shows immediate value.
2. While you live there, lay the foundation for the future. (Work on the laws, ethics, and big data rules).

In a nutshell:
The people who actually work in cancer care are ready to use AI to make their jobs easier and help patients faster. They aren't scared of the technology itself; they are just waiting for the "boring" stuff (rules and safety nets) to catch up. The paper suggests we start small with the helpful tools while we slowly build the big safety structures around them.

Technical Summary

1. Problem Statement

While Artificial Intelligence (AI) holds significant promise for improving efficiency and outcomes in healthcare, its implementation in oncology remains complex due to the field's multidisciplinary nature and high coordination demands. Existing literature often explores general attitudes toward AI in medicine or focuses on specific barriers, but there is a scarcity of research that:

• Specifically addresses the unique intricacies of cancer care.
• Integrates perspectives from a diverse range of stakeholders (patients, clinicians, AI researchers, administrators, and IT professionals) simultaneously.
• Provides a structured, quantitative framework to prioritize implementation strategies based on both importance and feasibility.

The study aims to bridge this gap by identifying perceived benefits, concerns, and priorities to guide the responsible, evidence-based (EB) implementation of AI in cancer care.

2. Methodology

The study employed a participatory, mixed-methods approach utilizing Concept Mapping, a structured process that integrates qualitative idea generation with quantitative multivariate analysis.

• Study Context & Participants:
  • Workshop: Conducted at the 2024 British Columbia (BC) Cancer Summit (Vancouver, Canada).
  • Brainstorming: 48 participants (patients, families, clinicians, researchers, IT, administrators) generated 265 initial statements regarding AI concerns, benefits, and priorities.
  • Expert Panel: A purposively sampled panel of 13 experts (54% clinicians, 31% AI researchers, 15% admin/patient experts) performed the sorting and rating tasks.
• Concept Mapping Procedure (6 Stages):
  1. Preparation & Generation: Workshop discussions generated raw statements.
  2. Organization: Statements were deduplicated, randomized, and rephrased, resulting in a final set of 100 unique statements.
  3. Sorting: Panelists grouped statements into piles based on conceptual similarity (average 5 piles per person).
  4. Rating: Each statement was rated on a 5-point Likert scale for Importance (1–5) and Feasibility (1–5).
  5. Representation (Analysis):
     • Multidimensional Scaling (MDS): Used to create a 2D spatial map where proximity reflects conceptual similarity based on sorting frequency.
     • Hierarchical Cluster Analysis (Ward's Method): Applied to MDS coordinates to identify clusters and subclusters. Optimal numbers were determined via silhouette analysis and the elbow method.
  6. Interpretation & Utilization:
     • Pattern Match Analysis: Compared importance vs. feasibility ratings.
     • Go-Zone Analysis: Categorized subclusters into four quadrants (High-High, Low-High, High-Low, Low-Low) to identify immediate priorities vs. long-term investments.
     • Subgroup Analysis: Compared ratings between clinical and non-clinical professionals using Mann–Whitney U and t-tests.
• Tools: Analysis was performed in Python (v3.12) using custom scripts adapted from Bar & Mentch.

3. Key Results

• Statement Generation: 265 initial statements were refined to 100 final statements covering concerns (32.8%), benefits (27.2%), and research priorities (40.0%).
• Cluster Structure: Two primary clusters emerged:
  • Cluster 1: "Challenges and Safeguards for AI Implementation" (62 statements).
  • Cluster 2: "Clinical Benefits and Efficiency Gains" (38 statements).
• Quantitative Findings:
  • Significance: Cluster 2 (Benefits) had significantly higher mean ratings for both Importance ($M=3.50$ vs $3.14$, $P<.001$) and Feasibility ($M=3.62$ vs $2.99$, $P<.001$) compared to Cluster 1.
  • Effect Size: The difference in feasibility was particularly large (Cohen's $d = 1.59$).
  • Subgroup Differences: No significant difference was found between clinical and non-clinical professionals regarding importance. However, clinical professionals rated feasibility for Cluster 2 significantly higher ($P=.029$).
• Subcluster Analysis (8 Subclusters):
  • Cluster 1 Subclusters: Focused on immediate operational concerns, ethics/safety, regulatory/privacy, and quality assurance. These generally had lower feasibility ratings.
  • Cluster 2 Subclusters: Focused on research/knowledge, direct workflow improvements, clinical quality, and patient communication. Subcluster 2.2 (Direct workflow improvements) showed the highest ratings for both importance and feasibility.
• Go-Zone Analysis:
  • Quadrant I (High Importance/High Feasibility): 36 statements (36%). These represent "Priority" items, primarily workflow-aligned tools (e.g., documentation support, triage).
  • Quadrant III (High Importance/Low Feasibility): 13 statements (13%). These represent "Strategic" items requiring R&D and policy work (e.g., data governance, algorithmic transparency).

4. Key Contributions

1. First Concept Mapping Study in Oncology AI: This is the first study to use concept mapping to structurally organize stakeholder perspectives specifically for AI in cancer care, moving beyond thematic reviews to quantitative spatial analysis.
2. Differentiation of Implementation Layers: The study empirically distinguishes between workflow-aligned applications (perceived as ready for immediate implementation) and system-level governance requirements (perceived as critical but requiring long-term investment).
3. Actionable Framework: It provides a "Go-Zone" framework that allows healthcare organizations to sequence implementation efforts:
   • Short-term: Prioritize tools that integrate into existing workflows (high feasibility).
   • Long-term: Invest in infrastructure, policy, and governance (high importance, low feasibility).
4. Stakeholder Consensus: Demonstrated that diverse stakeholders (clinicians vs. non-clinicians) largely agree on the importance of AI benefits, though clinicians perceive the feasibility of implementation slightly higher.

5. Significance

The findings offer a practical blueprint for healthcare administrators, policymakers, and implementation scientists. By validating that stakeholders view workflow-integrated AI as a "quick win" while recognizing governance as a necessary but difficult long-term hurdle, the study suggests a phased implementation strategy.

This approach mitigates the risk of "analysis paralysis" by allowing organizations to deploy high-feasibility tools to build trust and demonstrate value, while simultaneously initiating the complex, multi-year work required for regulatory, ethical, and data governance frameworks. The study aligns with implementation science frameworks (like CFIR), confirming that intervention characteristics (workflow compatibility) and inner settings (readiness) are currently more favorable than outer setting factors (regulatory environment).

Limitations: The study is geographically limited to British Columbia, Canada, and relies on a purposively sampled expert panel (n=13), which may limit generalizability. However, the methodology is robust for generating actionable local insights and a transferable conceptual framework.