Audience: Anyone deciding whether the Algorithm Viewer is the right tool, or wanting to understand its purpose and scope.
The problem
Health risk prediction models are usually shared as equations, coefficient tables, or code. That form is precise but hard to understand: it is difficult to see, at a glance, how a predictor such as age or BMI actually shifts the predicted outcome, how two models differ, or whether a large relative effect corresponds to a large absolute one. Reviewers, collaborators, clinicians, and the researchers building the models all need a way to see a model’s behaviour, not just read its parameters.
The Algorithm Viewer exists to make that behaviour visible and interactive.
What it is
The Algorithm Viewer is an R Shiny application, developed by Project Big Life at The Ottawa Hospital, for visualizing health risk prediction algorithms. It plots a model’s behaviour as interactive curves — odds ratio, predicted risk, and relative risk — and lets you adjust a baseline “reference patient” and compare models side by side. It works with any algorithm expressed in the Model Parameters format developed by Big Life Lab.
The goal is to turn a static model specification into something a researcher or clinician can explore directly: change a predictor, watch the curve move, compare two models, and build intuition about what the model does.
Who it is for
- Researchers and statisticians building risk-prediction models, who want to sanity-check behaviour during development — verifying that curves move in the expected direction and spotting unexpected patterns.
- Clinicians reviewing a model, who want to understand its implications without reading its mathematics.
- Readers of a publication, who can be given a link to explore the exact algorithm a paper describes (see Including the Algorithm Viewer in publications).
- Students learning how statistical prediction models behave.
What it is good for
- Seeing how a single predictor moves the outcome across its whole range.
- Distinguishing relative effects (odds ratio, relative risk) from absolute effects (predicted risk) — a distinction that matters clinically and is easy to lose in a coefficient table.
- Comparing related models (for example sex-stratified models) on the same axes.
- Exploring “what-if” and personal-profile comparisons through the Me vs Ref tab.
- Sharing a model interactively without asking the audience to install anything (when deployed as a web app).
What it is not for
- It is not a modelling or fitting tool. The viewer visualizes models that already exist; it does not estimate, fit, or validate them. The statistics come from the underlying Model Parameters pipeline, not from the viewer.
- It is not a clinical decision-making tool. The plots are for exploration and understanding. Individual risk estimates should not be used to make care decisions unless the underlying algorithm has been validated and approved for that purpose.
- It does not read algorithms directly from the internet. The files it visualizes must be available on the local filesystem (see View your own algorithms).
- It only understands the Model Parameters format. An algorithm must be expressed in that format to be viewable.
Features at a glance
- Multiple plot types — odds ratio, relative risk, and predicted risk, for both continuous and categorical predictors, plus visualization of predictor interactions.
- Model comparison — load and compare multiple models side by side, each in a distinct colour, including sex-stratified or otherwise stratified models.
- Reference group configuration — adjust the baseline predictor values interactively to re-centre every plot on a reference patient of your choosing.
- Personal-profile comparison — the Me vs Ref tab compares a personal profile against a reference and breaks down which predictors drive the difference.
- Flexible deployment — run it locally in R, as a Docker container, via ShinyProxy for multi-user access, or as a browser-only Shinylive build.
Where it is heading
The Algorithm Viewer’s deployment roadmap envisions three complementary modes, rolled out in phases:
- Local development tool — the installable R package and Docker image (available today), letting scientists view their models with a single command.
- Public web application — a hosted deployment where users can upload and explore their own algorithms without installing anything.
- Hosted algorithm showcase — a mode where uploads are disabled and specific algorithms are shared for viewing, linkable by URL — designed for publications and dissemination.
These options are not mutually exclusive; they target different audiences and can run in parallel. The design details and open questions are discussed in the project’s deployment specification.
Next steps
- What is Model Parameters? — the format the viewer depends on.
- Viewing the Built-in HTNPoRT Algorithm — see the viewer in action.