NeuroRiskX
A stroke-risk model that shows its work. Every prediction ships with a per-patient SHAP explanation, the decision threshold it was scored against, and the model's own held-out metrics, so nothing on the screen is taken on trust.
Try the live demoWhat it does
The challenge
A risk score on its own is unusable in any setting where someone has to act on it. If a model says 72 out of 100 and cannot say why, a professional has no basis to agree or disagree with it, and no way to catch it when it is wrong.
Our approach
- 1
Explanation as a first-class output
Every scored assessment returns a per-patient SHAP breakdown, so the answer to 'why this score' is part of the response rather than a separate analysis.
- 2
The threshold made visible
The score is shown against the decision threshold it was actually judged against, so the flag is legible rather than a black-box verdict.
- 3
Model facts reported, never typed
The held-out ROC-AUC, recall, and precision on screen are reported by the scoring service itself, so the page cannot drift from the model it describes.
- 4
A what-if panel over the live model
Inputs can be changed and rescored against the real model, turning a static score into something a professional can interrogate.
The outcome
A working explainable-ML demo: a real patient from the held-out test set is scored by the live model, flagged against its threshold, and explained feature by feature, with no account required to see it.
What you get
- A model whose every output is explainable
- An interface a non-specialist can interrogate
- Honest metrics, reported rather than claimed
From the build
Have a project like this in mind?
Tell us what you're trying to build. We'll scope it with you in a 30-minute call.