✳ StrokeLensRESEARCH DEMO

MACHINE LEARNING · HEALTHCARE · EXPLAINABLE AI

A prediction is only
half the story.

Understanding why matters.
Explore stroke prediction models and the features behind their decisions.

Built by Richard Huang / UCL · COMP0172

EXPERIMENT SNAPSHOT● NOTEBOOK RESULTS
0.8484ROC AUC

Random Forest · max depth 5

88.71%Stroke recall
17.46%Stroke precision

High sensitivity, with a substantial false-positive tradeoff.

5,110Original dataset records
4Model families
1,022Held-out test records
10Input features

01 / BENCHMARK

Compare the models.

One held-out test set. Five configurations. See the tradeoffs beyond accuracy.

Download comparison chart ↓

Model performance

0.000.250.500.751.00

Exact saved notebook outputs · Cell 29 · Higher is better

Inspect all recorded metrics
ModelAUCRecallPrecisionF1Accuracy
Download metrics CSV ↓

02 / UNDER THE HOOD

Make the model legible.

Original SHAP visualizations, extracted directly from the project notebook.

EXPLAINABLE AI
Original SHAP mean absolute feature importance plot from notebook cell 34

Mean absolute SHAP values rank features by average contribution magnitude.

03 / RESEARCH CONTEXT

From data to explanation.

A transparent view of the original experiment, including its limitations.

01

Prepare

Drop ID and the single “Other” gender row; 5,109 records remain. Split 80/20 with random state 42. Fit mean BMI imputation on training data.

02

Train & compare

Classifier pipelines use one-hot encoding, numerical scaling and RandomOverSampler. The notebook also contains a separate SMOTE experiment.

03

Explain & qualify

Inspect original SHAP feature importance, beeswarm and age dependence plots. Saved outputs are presented as historical experiment results.

A portfolio research demo.

This demo presents saved research results; it does not run live patient predictions. The notebook contains execution-state inconsistencies and no exported fitted model. Reproducible retraining, calibration, subgroup evaluation and external validation would be needed before considering clinical use. Not a medical device or diagnostic tool.

LET’S BUILD SOMETHING EXPLAINABLE

Turn your data into
decisions you can understand.

Explore my work in machine learning, healthcare AI and data science.

Visit Richard’s portfolio ↗