Table 4 Pre- and post-launch timeline for RAI Governance at MSK
From: Responsible Artificial Intelligence governance in oncology
Timepoint | Activities | Outputs |
|---|---|---|
T-12 months & prior | • Ethics Committee workgroup • Many clinical, operational and research AI projects • LLMs and GenAI become GA in the market | • 6 Ethical AI principles ratified (beneficence, safety, transparency, user engagement and agency (including patients), stewardship, and mitigation of bias) • Many AI-related publications by “AI Practitioners” at MSK • AI/ML engineering team initiated, including Sandboxes with LLMs for “approved use” |
T-9 months | AI Task Force (AITF) formed Members: Leadership from strategy, research, computational and molecular oncology, clinical services, operations, ethics, informatics and technology | • AI program model (see Fig. 1) • Oncology AI goals and priorities defined • 7 subgroups formed to advise AITF (basic research; translational research; clinical care & research; operations; external partnerships; governance & operating model; data and tech foundation) |
T-6 months | AI Task Force Strategic Planning | • AI vendor partnership model (see Fig. 1) • AI project inventory completed (80 programs over 9 programmatic domains) |
T-3 months | AI Task Force Final Report | • Strategic funding priorities ratified • Minimum core capabilities and major barriers identified • Board of Trustees approval of 5-year roadmap • Incorporation into MSK’s 2030 strategic plan • AI “governing body” proposed |
Time 0 (January 2024) | AI Governance Committee (AIGC) formed Members: Key stakeholders from ethics, clinical care, research, quality and safety, legal, compliance, strategy, communications, nursing, informatics, and technology | •Version 1 charter released • Model Information Sheet drafted (aka “Nutrition Label” or “Model Card”) • Definitions of “in-scope/to-be-governed” AI/ML and other related tools released, focused on clinical and operations models • Updated request intake and “idea refinement” process ratified for new AI project requests (using “AI” tag in our ticketing system) |
T + 3 months | AIGC processes: AI intake and model management; retrospective model reviews | • Novel Lifecycle Management model drafted (adapted from Duke ACBDS model36,37) • Novel application of Risk Model ratified (See Table 1; adapted from UC Berkeley risk model36) • Nomograms on public website reviewed and renewed |
T + 6 months | AIGC Model Lifecycle Management | • 2 workgroups formed (lifecycle governance; regulatory) • Novel Lifecycle Management model (iLEAP) ratified (see Fig. 2, Panel a) • Model Information Sheet questions ratified (see Table 1) • Model Registry prototype developed |
T + 9 months | • AIGC model registration • Update policies • Integration with quality & safety processes • Initial education and comms planning | • 2 workgroups added (policy; quality & safety) • Model Registry go-live in production |
Current quarter | • Year-in-review for continuous improvement • Anticipatory interception of ai-induced adverse events | • Version 2 charter revisions in progress • Prototyping novel AI-induced adverse events (aiAE’s) capture and remediation process with Quality & Safety |
