Results of the AI Governance Approach: Frameworks, Tools, and Implementation Outcomes

The results section outlines the frameworks, processes, and tools developed during the design and implementation phases of our study, emphasizing their role in the governance of artificial intelligence in healthcare.

Transforming Healthcare with AI: Insights from Our Study

In recent years, the integration of artificial intelligence (AI) into healthcare has marked a revolutionary shift, with organizations racing to harness its potential. Our ongoing study aimed to design and develop frameworks, processes, and tools that not only facilitate AI integration but also define how we govern its application in clinical settings. This post dives into the results of the initial design and development phase, showcasing our journey to manage AI models effectively and responsibly.

Reusable Frameworks and Tools

From our study’s design and development phase emerged several reusable frameworks and tools that were instrumental during the implementation phase. These tools enabled the successful registration, evaluation, and monitoring of 26 AI models—including large language models—and 2 ambient AI pilots. We also conducted a retrospective evaluation of 33 clinical nomograms.

Nomograms serve as vital clinical decision aids, presenting numeric or graphical predictions based on various prognostic variables. They exemplify the importance of meticulous evaluation, which lays the groundwork for Responsible AI (RAI) governance in action.

Phase 1: Design & Development

AI Task Force Results

The AI Task Force (AITF) was pivotal in shaping our AI program framework. Their efforts identified four main challenges that must be addressed to optimize AI model development:

1. High-Quality Data: The backbone of any effective AI model.
2. High-Performance Computing: Essential for processing massive datasets.
3. Talent Capacity: Skilled professionals are crucial for navigating AI complexities.
4. Policies and Procedures: Clear guidance is necessary to ensure compliance and alignment with organizational goals.

By Q4 2023, the AITF had compiled an AI Model Inventory that highlighted 87 active projects across various domains. Notably, the leading development teams included Strategy & Innovation, Medical Physics, and Computational Pathology, emphasizing the collaborative nature of our AI initiatives.

In response to identified challenges, the AITF prioritized future investments into five strategic goals, advocating for enhanced high-quality data curation via AI-enabled processes.

AI Governance Committee Results

Our AI Governance Committee (AIGC) developed the iLEAP model, which stands for Legal, Ethics, Adoption, and Performance. It introduces decision gates that ensure a structured approach to AI model management. This model emphasizes:

• Three paths for AI practitioners: Research, Home-Grown Build, and Acquired/Purchased.
• A collaborative review process fostering scientific freedom while ensuring safety and quality.

One of the key tools developed was the Model Information Sheet (MIS), which helps keep track of models in various stages and captures anticipated adverse events (aiAEs). The AIGC also devised a risk assessment model for evaluating AI models, enabling quantifiable metrics for aiAEs, risk factors, and mitigation strategies.

Phase 2: Implementation

AI Model Portfolio Management

In the implementation phase, the AIGC manages the AI model portfolio as a dynamic collection, allowing for real-time tracking and evaluation. The Model Registry plays a critical role in tracking models through the iLEAP gates, informing iterative refinements to the model deployment process.

Notably, our analysis showed a 63% increase in AI project intake from 2023 to 2024, underscoring growing demand and interest.

Case Studies of RAI Governance in Action

To illustrate our governance framework’s application, we present two case studies of AI models in action: one acquired and one developed in-house.

1. Radiology FDA-Approved AI Model: A third-party vendor created this AI model to assist in breast cancer detection through image analysis. Following rigorous assessments and the establishment of a quality assurance program, the AIGC approved its deployment, heralding improved outcomes in radiological assessments.

2. In-House Developed Tumor Segmentation Model: Developed by our Medical Physics department, this model offers new capabilities in tumor segmentation using MRI images. The AIGC’s risk assessment deemed it medium risk, yet the established quality assurance framework enabled its approval for clinical use.

Conclusion

The frameworks, models, and governance processes developed through our study position us at the forefront of responsible AI use in healthcare. They provide a blueprint for not only improving patient outcomes but ensuring ethical, safe, and effective AI integration into medical practice. As we continue to grow our AI initiatives, these reusable tools will guide our exploration into the future of healthcare innovation, potentially transforming the way we deliver care.

By embedding RAI governance into our operational framework, we strive to lead the charge in harnessing the promise of AI—with the ultimate goal of enhancing quality of care and patient safety.