RheocorAI v3.0
Open the dashboard

Know your heart-disease risk, and the reasons behind it.

RheocorAI estimates the chance that one of your major heart arteries is significantly narrowed, and shows the clinical factors behind that estimate along with how much it varies across models.

You enter 13 routine clinical numbers, such as age, blood pressure, cholesterol, and ECG and stress-test results, and four machine-learning models each estimate the probability of obstructive coronary artery disease, meaning a major heart artery narrowed by 50% or more. The models were trained on the UCI Cleveland Heart Disease dataset and checked with 5-fold cross-validation, bootstrap confidence intervals, and DeLong tests.

Open the dashboard How it works

Research and educational tool. Not a medical device, and not a substitute for clinical evaluation. Inputs are sent to the model server only to compute a score, and are never stored.

Illustrative example: a 61-year-old with exertional chest pain
0%
–
Probability of obstructive coronary artery disease
XGBoost–
Random Forest–
Logistic Regression–
Neural Network–
thal
cp
oldpeak
thalach
Mock values for illustration. The live dashboard computes these from your inputs.
Re-scores as you adjust inputs No accounts · nothing stored

The Dashboard

Type in a few clinical numbers and watch the score, the four models, and the reasons update as you go.

127.0.0.1:5000/app
The RheocorAI dashboard: a risk gauge, all four model probabilities, the top reasons behind the score, and a clinical-context panel.
  1. 1
    A clear probability

    A score from 0 to 100% with a color band, plus a range that shows how much it moves across the cross-validation models.

  2. 2
    All four models together

    Logistic regression, random forest, XGBoost, and a neural network each report their own number, next to the average and how much they agree.

  3. 3
    The reasons behind it

    SHAP shows which of the entered inputs pushed the score up (red) or down (blue).

  4. 4
    Clinical context

    A Framingham-style ten-year comparator and an age- and sex-matched percentile sit beside the score.

Try it with your own numbers

Why it shows its work

Most heart-risk calculators hand you a single number and stop there. You never find out which of your inputs mattered, or how sure the tool actually is. I wanted the opposite.

RheocorAI runs four different models on the same inputs and shows all of them, so you can watch them agree on an easy case and pull apart on a hard one. For every score it lists the exact inputs that pushed it up or down, and it shows how much the answer would move if the models had trained on a slightly different group of patients, so a shaky estimate looks shaky on screen. What you see is what the models produced, with nothing smoothed over to look more confident.

Whatever you type is used to compute a score and then forgotten. No account, no tracking. And if you want to check any of it, there is a model card, a written paper, and a pipeline that rebuilds every number with one command.

Development Log

Click to read Release Notes.

Latest release v3.0 (June 2026)
January 2025

Data pipeline

Built the ingestion and cleaning pipeline for the UCI Cleveland dataset: missing-value handling, target binarization, stratified splitting, and standardization.

February – March 2025

Model selection

Settled on four architectures spanning the bias–variance spectrum (Logistic Regression, Random Forest, XGBoost, and a small neural network), chosen to make the complexity-versus-accuracy question testable.

April – May 2025

v1.0 First prototype

A working binary classifier with a basic input form. Logistic regression only, no explainability, but live end to end.

June 2025

v1.4 SHAP explainability

Every prediction now ships with its top contributing features, the explainability layer that became the project's core.

July 2025

v1.8 Dashboard redesign

Dark sidebar, risk gauge, live slider updates, responsive layout, keyboard shortcuts, and reduced-motion support.

August 2025

v2.0 Four models

Random Forest, XGBoost, and the neural network joined the baseline, with one-click switching, an ensemble mean, and side-by-side comparison.

September 2025

v2.2 Cross-validation cohort

5-fold stratified cross-validation for every architecture, with the fold models reused at prediction time to show a per-patient stability range.

October 2025

v2.4 Clinical context

Added the Framingham-family comparator and the age- and sex-matched population percentile panel.

November 2025

v2.6 Resources & care

Map-based facility finder (OpenStreetMap), region-aware emergency numbers, and country-specific heart-health resources.

December 2025 – April 2026

v2.8 Product surface

Landing page, model card with citation generator, PDF export, patient history, What-If simulator, federated research view, and onboarding tour.

June 2026

v3.0 The rigor release

Full methodology audit: leak-free training and per-fold preprocessing, bootstrap CIs and DeLong tests, calibration analysis, unmodified model outputs with a separate guideline advisory, and every public claim re-derived from the committed results.

If you want to look closer

The app has a small research corner, all of it computed on the patient you have entered. You can watch a model learn across simulated hospitals that never share their data, compare three ways of explaining the same score, and see how the estimate shifts as the patient gets older.

Open the Research view

Try it now, with full privacy.

No account, no tracking, and nothing you enter is stored.

Open the dashboard