Moving towards vertically integrated artificial intelligence development.

Area of Science:

• Clinical Artificial Intelligence (AI)
• Health Informatics
• Biomedical Engineering

Background:

• Despite significant investment, clinical AI research translation into deployed solutions remains limited.
• Current research often adopts a model-centric approach, focusing on algorithm optimization and performance on available datasets.
• Challenges include real-world unpredictability, out-of-dataset scenarios, deployment infrastructure, and lack of clear clinical value.

Purpose of the Study:

• To propose a vertically integrated approach for clinical AI development.
• To address the gap between AI research and real-world clinical application.
• To provide practical recommendations for increasing the translational potential of clinical AI.

Main Methods:

• Defined a vertically integrated approach encompassing impact evaluation, data lifecycles, and AI production.
• Examined implementation in two contrasting AI development pipelines: an "AI factory" and a cervical cancer screening platform.
• Reviewed challenges and future approaches, including federated architectures.

Main Results:

• A vertically integrated approach, incorporating cross-disciplinary considerations early in development, is crucial for successful AI translation.
• Implementation in diverse settings (e.g., Mayo Clinic, Paps AI) demonstrates the adaptability of this approach.
• Practical recommendations and insights into future challenges were provided.

Conclusions:

• Vertically integrated teams and development practices can significantly increase the translational potential of clinical AI projects.
• Moving beyond a model-centric view is essential for robust AI deployment in healthcare.
• Future clinical AI growth necessitates integrated development strategies for greater real-world impact.