Related Experiment Videos
Development and Pilot Evaluation of a Standardized and Accessible Rubric (OMNI) for Assessing Generative AI in Health
Erik Cornelisse1, Dib Kousa1, Bradley D Menz1
1College of Medicine and Public Health, Flinders Health and Medical Research Institute, Flinders University, Flinders University Health and Medical Research Building, Sports Rd, Bedford Park, Adelaide, South Australia, 5042, Australia, +61 8 8201 5647.
Background:
The rapid adoption of generative AI in the community and health care necessitates reliable, human-led evaluation.
Objective:
This study aimed to develop a standardized rubric to support health care professionals, administrators, and researchers in evaluating generative AI outputs in health contexts.
Methods:
A rubric was developed through a 4-stage mixed methods study: literature synthesis, expert refinement, pilot application, and post-pilot feedback. The rubric was informed by 69 publications and refined through iterative feedback from a multidisciplinary expert panel (n=23). The pilot rubric was applied to 4 health care scenarios: responses to health-related Reddit questions, generation of health-related images and videos, responses to health professional continuing education questions, and automated physician-to-physician clinical documentation generation. Via these scenarios, 15 clinicians (medical practitioners, pharmacists, and medical oncologists) evaluated 250 input-output pairs using domain-level ratings and qualitative comments. The evaluated outputs included 200 AI-generated responses across the 4 health care scenarios and 50 physician-tagged replies in the health-related Reddit questions scenario.
Results:
The final rubric, OMNI, comprised 10 evaluation domains (Accuracy, Safety, Recognizes Uncertainty, Completeness, Communicates Clearly, Empathetic, Context Aware, Referencing, Follows Instructions, and Avoids Coerciveness) assessed on 4-point Likert scales, supported by detailed instructional and example materials. Via the 4 pilot scenarios, assessors generated 9600 domain-level ratings and 4010 qualitative comments. The mean pairwise agreement for the 4-point Likert groups ranged from 30.8% to 53.4%, increasing to 59.3% to 79.6% when collapsed into broad agreement categories; Gwet's AC2 ranged from 0.49 to 0.80. OMNI differentiated strengths and weaknesses between models within scenarios, while recurrent limitations in Referencing, Recognizes Uncertainty, Completeness, Empathetic, and Accuracy were identified across scenarios. Multimodal image and video outputs had the highest occurrence of negative rating domains. Assessors reported OMNI was usable and indicated they would reuse it.
Conclusions:
OMNI is a standardized, accessible rubric for expert-led evaluation of generative AI outputs in health contexts. Pilot application demonstrated that OMNI could differentiate AI performance across models and domains while generating structured feedback. Further work is needed to assess its applicability to purpose-built health AI tools, agentic systems, and organizational evaluation settings.