AI EVALUATION IN HEALTHCARE: A Decision-Analytic Tool for Selecting Meaningful Performance Metrics

STEP 1 — START HERE

What is your evaluand?

What is the ‘thing’ you are evaluating?

→  AI model 

·       Sensitivity and specificity metrics[2]

·       Explainability metrics (e.g., heatmaps, feature importance scores)[3]

·       Traceability metrics (e.g., logging of AI inputs/outputs, audit trail)[4]

·       Bias and fairness metrics (e.g., statistical parity, equalised odds)[5]

·       Other relevant metrics depending on the  stage of AI lifecycle (design, development, evaluation,  or post-deployment)

→  Implementation  

·       Cost

·       Fidelity[6]

·       Usability and acceptability (e.g., System Usability Scale score, time to clinician competency)[7]

·       Other implementation-specific metrics based on taxonomies of implementation science constructs

→  Impact   → Meaningful benefit to people or organizations

·       For patients (e.g., improved health outcomes)

·       For clinicians (e.g., minutes saved per encounter/day, confidence regarding clinical decision-making)

STEP 2

What is the unit of analysis for your evaluand?

The unit of analysis determines the level at which metrics are collected, aggregated, and reported. The same AI system assessed at project level will use different metrics than when assessed across a multi-country partnership. These are some potential units:

·       AI Model

·       An individual AI Project  

·       Program

·       Alliance, Network or International Partnership 

STEP 3

What is your evaluation lens?

There are many different types of evaluations beyond randomized controlled trials. You will need different metrics depending on what evaluation lens you use. These are not just ‘nice-to-have’ domains you should be aware of. These are meaningful metrics that you can operationalize, track, and hold yourself and your project accountable for. For example:

→  Theory-based Evaluation   → contextual indicators tracked over time, such as staff attitudes before AI introduction (baseline), existing challenges with electronic health records into which your AI tools are being integrated, team assumptions about the project, and the regulatory environment [8]

→  Feminist Evaluation  → % budget for patient engagement; childcare costs covered (Y/N)

→  Made in Africa Evaluation   → % salary spent in-country; number and composition of local team

→  Indigenous Evaluation   → Indigenous data sovereignty principles in place (Y/N)[9]

→  Blue Marble Evaluation   → Energy consumption per AI task; use of model pruning, quantisation, edge computing (Y/N)[10]