AI-assisted screening tools for diabetic retinopathy (DR) are increasingly being considered for deployment within national health systems. In Aotearoa New Zealand, the diabetic retinal screening programme serves over 200,000 people, and AI tools — including THEIA, a system prospectively validated within the NZ programme — are being evaluated for clinical integration. Published clinical trials consistently report strong aggregate diagnostic accuracy. But aggregate performance metrics can obscure something critical: does the tool perform equally well for everyone?
This talk presents a simulation-based evaluation framework, built entirely in Python, that interrogates the equity implications of deploying AI screening tools across demographically diverse populations. Māori and Pacific peoples in New Zealand experience higher rates of both diabetes and diabetic retinopathy, greater socioeconomic deprivation, and lower screening attendance than NZ European populations. If an AI tool performs less accurately for these groups — even modestly — deployment at scale could widen existing health disparities rather than close them.
The Framework
The framework generates a synthetic cohort of 10,000 patients calibrated to the NZ diabetic screening population, with demographics (ethnicity, age, sex, NZDep deprivation quintile) and true disease status drawn from published epidemiological data. Three AI tool profiles are then simulated: the IDx-DR system (from its pivotal trial and a recent meta-analysis) and the NZ-developed THEIA system. Each tool is evaluated under two scenarios — equal performance across subgroups, and differential performance informed by bias patterns documented in the clinical AI literature.
Performance metrics (sensitivity, specificity, PPV, ROC curves) are disaggregated by ethnicity, area deprivation (NZDep), and their intersection, with statistical testing to identify significant disparities.
All simulation parameters are fully documented with citations in a
config/parameters.yaml file, making the framework transparent, reproducible,
and adaptable. The code is structured as three Jupyter notebooks (cohort
generation, diagnostic accuracy evaluation, equity analysis) backed by a
clean Python module layer — so anyone can swap in their own AI tool's
performance data and re-run the analysis for their own health system context.
Key Findings
Even modest subgroup-level accuracy gaps — easily obscured in aggregate trial reports — translate into meaningfully worse outcomes for already-disadvantaged populations when projected across a national programme. This is not a hypothetical concern. Studies of AI diagnostic tools in chest radiology, ophthalmology, and dermatology have documented precisely this pattern.
Why Python, Why This Matters
Python provides all the tools needed to build rigorous equity evaluations: NumPy and Pandas for cohort simulation, SciPy for statistical testing, and Matplotlib/Seaborn for disaggregated visualisation. This talk demonstrates that pre-deployment equity evaluation doesn't require proprietary software or large clinical datasets — just well-structured Python and publicly available literature.
Attendees will leave with: an understanding of why aggregate AI performance metrics are insufficient for equitable deployment decisions; a mental model for simulation-based pre-deployment evaluation; and an open-source Python framework (MIT licence) they can adapt to their own context.
This talk is for Python practitioners interested in data science, healthcare, and responsible AI. No clinical background is needed — the focus is on methodology and Python implementation, accessible to intermediate Python users and above.
Talk outline (~30 min):
I am a computational epidemiologist and AI researcher based in Tāmaki Makaurau Auckland, Aotearoa New Zealand. My PhD is in computational cosmology, but I have spent the last several years building models at the intersection of infectious disease, health data science, and responsible AI — with Python as the primary tool for everything from ETL pipelines to ensemble forecasting.
Lately, I have been increasingly interested in how AI tools get deployed in health systems — and specifically, whether the people who need them most are actually well-served by them. That question is what drove me to build the equity evaluation framework I am presenting at this conference.
I care a lot about open, reproducible research: the code for this talk is MIT-licensed and designed to be adapted. If you work in health, data science, or responsible AI — or you are just curious why aggregate accuracy metrics can be quietly dangerous —
I hope this talk is useful to you.
