Breast cancer · AI decision support

Estimating breast cancer risk with an artificial neural network.

A research-grade prototype that takes nine accessible clinical and demographic predictors — age, BMI, glucose, insulin, HOMA, leptin, adiponectin, resistin, and MCP-1 — and returns a calibrated risk estimate. Trained on the UCI Breast Cancer Coimbra dataset and designed to support, not replace, qualified medical judgment.

Sign in to begin Learn how it works
Dataset
UCI Coimbra · 116 records
Model
ANN · 2 hidden layers
Validation
Stratified 5-fold CV
Standard
TRIPOD+AI aligned
Capabilities

A complete prediction workflow, end to end.

Beyond a static research script — every part of the pipeline, from data entry to printable report, is built into a coherent application.

Guided data entry

Nine clinical and demographic inputs with realistic ranges, unit hints, and real-time validation to minimise entry error.

ANN inference

A trained multi-layer perceptron returns a calibrated probability and three-tier risk classification within a fraction of a second.

Visual results

Each prediction is shown with a probability gauge, confidence score, and an at-a-glance health advisory note.

Assessment history

Every assessment is saved, searchable, and reviewable with the inputs, prediction, and timestamp preserved.

Model evaluation

Built-in dashboard for accuracy, precision, recall, F1, ROC-AUC, and calibration — compared against logistic regression and random forest.

Printable report

Each assessment can be downloaded as a clean, print-ready report with the disclaimer and methodology summary attached.

Workflow

From input to interpretation in four steps.

Mirroring the Chapter 3 activity diagram, every prediction follows the same transparent pipeline.

Enter predictors

Nine fields covering demographics, anthropometry, metabolic, and inflammatory biomarkers.

Preprocess

Server-side validation, z-score standardisation, feature ordering, and consistency checks before inference.

ANN inference

The trained network produces a probability between 0 and 1 for the positive class.

Interpret

Probability is mapped to Low, Moderate, or High risk and presented with charts and an advisory note.

Methodology

A conservative architecture for a small clinical dataset.

The model is a multi-layer perceptron with two hidden layers using ReLU activation, sigmoid output, binary cross-entropy loss, and Adam optimisation. Dropout and early stopping guard against overfitting — essential on a 116-record dataset. Training uses an 80:20 stratified split with 5-fold cross validation inside the development set.

The pipeline emphasises calibration alongside discrimination, reports performance against logistic regression and random forest baselines, and follows TRIPOD+AI reporting guidance.

9
Clinical predictors
116
Training records
2
Hidden layers
5×
Stratified CV folds

ANN architecture

Keras · TensorFlow
Input 9 predictors Hidden 1 ReLU Hidden 2 ReLU + Dropout Output Sigmoid
Research grounding

Built on the published evidence.

Selected studies that informed the architecture, dataset, and validation strategy.

2018

Patrício et al.

Using resistin, glucose, age and BMI to predict the presence of breast cancer.

Why it matters — demonstrated that simple blood and anthropometric variables can support breast cancer prediction. The basis for the Coimbra dataset used in this project.
2019 · 2022

Yala et al.

Deep learning for breast cancer risk prediction (Mirai), with multi-institutional validation.

Why it matters — a landmark series showing neural-network-based risk models can outperform classical statistical tools and generalise across diverse populations when carefully validated.
2021

Macaulay et al.

Breast cancer risk prediction in African women using random forest classifier.

Why it matters — argues for population-aware models in African contexts. Motivates the local adaptability emphasis of this prototype.

Ready to begin an assessment?

Sign in to enter clinical predictors and receive a risk estimate.

Sign in to begin