Machine learning · Public health

Equitable vaccine
distribution.

A two-level framework for deciding who should receive a scarce vaccine first—and how supply should move across public-health regions.

Context: COVID-19 vaccine rollout
Scope: Individuals + Ontario PHUs
Methods: Clustering + classification

The challenge

Scarcity turns allocation into a question of judgment.

A fair rollout has to balance health risk, exposure, population characteristics, and geography—without hiding consequential decisions inside an opaque model.

The project proposed a transparent decision-support system with two connected layers. The first assigns an individual Provincial Priority Level. The second clusters public-health units so supply can be distributed using comparable regional characteristics. Expert review remains part of both layers.

The model

One framework.
Two levels of allocation.

Individual level

Who goes first?

People with health data are grouped into ten clusters using unsupervised learning. Each cluster receives a Provincial Priority Level from 1 (highest) to 10 (lowest), based on age, underlying health conditions, income, and exposure-related factors.

Demographics and income
Health conditions and clinical indicators
Exposure, density, and essential work

Public-health-unit level

Where does supply go?

Ontario public-health units are clustered using census and COVID-era indicators. Allocation can then reflect the population and risk profile of each region instead of political boundaries alone.

Population age and household composition
Active cases, resolved cases, and deaths
Occupation, income, housing, and health services

Decision flow

From data to an appointment queue.

01
Assess
Use available demographic, clinical, and exposure data to group people with similar profiles.
02
Assign
Translate cluster characteristics into a transparent priority level, with expert oversight.
03
Predict
For people without existing health data, use a registration questionnaire and a supervised classifier to estimate the same priority level.
04
Schedule
Place registrants into a location-specific queue and offer the next appointment to the highest-priority person.

Model evidence

The working analysis.

Individual-level vaccine prioritization model using unsupervised and supervised learning — Individual-level feature design and classifier comparison.

Ten population clusters with assigned Provincial Priority Levels — Ten clusters translated into an explicit vaccination order.

Public-health-unit clustering model for vaccine distribution — Regional allocation using census and public-health indicators.

Appointment queue example ordered by Provincial Priority Level — A location-level appointment queue ordered by priority.

AI ethics

Fairness is part of the architecture.

The framework treats transparency, confidentiality, security, and bias mitigation as product requirements—not a postscript.

Keep decisions legible. Document how priority levels are produced and avoid an unexplained black box.
Protect patient data. Secure collection, transformation, and storage, with clear consent and intended-use boundaries.
Audit inputs and outcomes. Use representative training data and inspect how sensitive variables could encode bias.
Preserve human review. Let an expert panel examine clusters and individual characteristics before priority levels become policy.

Original deck