Comparing a Top-down and a Bottom-up Approach for Implementing AI

The implementation of artificial intelligence (AI) in radiology has been slow despite the increasing availability of AI tools designed to assist radiologists in image diagnostics [1,2,3]. In Norway, large-scale AI implementation in healthcare is a new endeavor, and limited knowledge exists about effective strategies for such processes [4]. To address this, a formative evaluation was conducted on the implementation strategies employed by two Norwegian health regions—Health Region 1 (H1) and Health Region 2 (H2), both of which began their AI efforts in 2020. These regions adopted contrasting approaches, allowing for a comparative analysis of their benefits and limitations, particularly concerning temporal and co-constructional aspects of implementation [5,6].

H1 adopted a bottom-up, innovation-driven strategy to implement AI solutions quickly and gain practical insights into the process. Their approach emphasized stakeholder involvement at every stage. From 2020 to 2022, they conducted a procurement process without a comprehensive strategy, relying instead on preconditions defined by the project group and radiologists. Implementation began in 2023 with deployed one single algorithm at one hospital. Each step was documented and evaluated before expanding to other hospitals. This "learning by doing" approach generated actionable insights, which were compiled into an AI starter kit available for use by other health trusts. Organizational change processes accompanied the implementation, including workshops to address workflow changes and extensive communication with stakeholders. H1’s strategy proved effective for deploying simpler algorithms, such as a fracture detection solution, due to its speed and lower resource demands. It enabled real-world testing, workflow adjustments, and the creation of knowledge from clinical use. However, this approach may lack the rigorous evidence required to build trust in more complex algorithms, particularly those influencing life-altering clinical decisions.

H2 followed a top-down, research-based strategy, emphasizing evidence generation and regional coordination. Led by a regional project group, their efforts resulted in the production of three comprehensive reports between 2021 and 2024. The first report introduced Norway’s first regional AI strategy, the second outlined a strategic framework for AI in radiology, and the third provided practical implementation guidance. These reports were collaboratively developed with input from all health trusts in H2, building a solid theoretical foundation. H2’s plan involved a region-wide rollout of AI solutions, with the same tools implemented sequentially across all health trusts. While this approach ensures thorough preparation and quality assurance—critical for complex algorithms with significant implications for patient care—it has been extremely time-intensive. As of 2024, no AI solutions have been implemented, raising concerns about the risk of algorithms becoming outdated due to the rapid pace of AI advancements.

Key Insights and Conclusion

H1’s innovative, bottom-up approach demonstrated the benefits of speed and adaptability for simpler AI tools, while H2’s research-driven, top-down strategy highlighted the importance of rigorous evidence for complex life altering algorithms. Neither approach alone appears sufficient for all AI implementations, suggesting a need for a balanced, co-constructed strategy that integrates the strengths of both. This study underscores the importance of mutual learning and flexibility to create more efficient future AI implementation strategies.

References

1. van Leeuwen KG, Schalekamp S, Rutten MJ, van Ginneken B, de Rooij M. Artificial intelligence in radiology: 100 commercially available products and their scientific evidence. Eur Radiol. juni 2021;31(6):3797–804.
2. Your Europe. CE marking – obtaining the certificate, EU requirements. Available on: https://europa.eu/youreurope/business/product-requirements/labels-markings/ce-marking/index_en.htm
3. Cresswell, K., et al., The need to strengthen the evaluation of the impact of Artificial Intelligence-based decision support systems on healthcare provision. Health policy, 2023. 136: p. 104889
4. Silsand, L, Kannelønning, Severinsen GH Ellingsen G. Enabling AI in Radiology: Evaluation of an AI Deployment Process. Studies in health technology and informatics, 2024. 316:580-584.
5. Karasti, H. Infrastructuring in Participatory Design. in 13th Participatory Design Conference 2014.
6. Hanseth, O. and Lyytinen K. Design Theory for Dynamic Complexity in Information Infrastructures: The Case of  Building Internet. Journal of Information Technology. 2010; 25(1): 1-19.