DOI: 10.1002/lrh2.70131 ISSN: 2379-6146
Generating Real‐World Evidence From
AI
Deployed in Health Systems
Lily Poursoltan, Aaron Boussina, Jie Cao, Eileen Kim, Poorva Satish Bedmutha, Hetvi Majmudar, Mattheus Ramsis, Alexander Khalessi, Karandeep Singh ABSTRACT
Background
Health systems increasingly deploy predictive and generative artificial intelligence (
AI
) to improve care delivery and outcomes. Predictive
AI
models estimate the risk of future events to inform targeted interventions, while generative
AI
systems extract, summarize, and communicate information and may automate selected tasks. Routine deployment generates clinical, operational, and interaction data that can support evaluation of the effectiveness and safety of
AI
‐enabled workflows.
Methods
We review approaches to generating real‐world evidence (
RWE
) from routine deployment of predictive and generative
AI
across four implementation configurations: silent deployment, human‐in‐the‐loop (
HITL
), human‐on‐the‐loop (
HOTL
), and automated. For each configuration, we describe the evaluation questions, routinely generated data, and feasible study designs.
Results
Each implementation configuration supports different forms of evidence about model performance, workflow effectiveness, human oversight, and safety. For both predictive and generative
AI
, silent deployment supports assessment of local performance and subgroup differences. Once predictive
AI
drives action, evaluating the effectiveness of an implemented workflow (
HITL
,
HOTL
, or automated) generally requires causal inference approaches. For generative
AI
,
HITL
implementations can generate data on edits, overrides, time on task, usability, cognitive burden, and safety‐relevant failure modes.
HOTL
and automated generative
AI
implementations can generate data on escalation performance, human follow‐up work, downstream outcomes, and safety signals.
Conclusion
AI
implementation should be paired with methods for generating
RWE
on the effectiveness and safety of
AI
‐enabled workflows. In
HOTL
and automated configurations, escalation performance and residual human workload become particularly important components of evaluation.