NeuroWeave Simulator : Decoding Deep Learning

NeuroWeave Simulator

Created by Dr. Sharad Maheshwari MD (imagingsimplified@gmail.com)
Institute For Responsible Healthcare AI (IRHAI)

Educational Tool

Bridging Healthcare
& Artificial Intelligence

NeuroWeave is an interactive, browser-based educational simulator designed to demystify how Deep Learning and Artificial Neural Networks are applied to real-world medical data.

How to Use the Simulator

1

Select a Clinical Scenario Choose a healthcare dataset. Each presents a different geometric pattern of data that the AI needs to learn.

2

Design the Architecture Adjust the number of Hidden Layers and neurons. Complex problems require deeper networks to solve.

3

Set Learning Rate Choose how aggressively the network updates. High learns fast but may overshoot; low is stable but slow.

4

Train & Observe Hit Play. Watch the Decision Boundary warp to separate the classes, and monitor the Loss (MSE) dropping.

The Clinical Scenarios

1. Chest X-Ray (Pneumonia)

Math: Linear Clusters

Context: AI processing pre-extracted X-ray metrics.

X: Lung Opacity. Y: Blood Inflammation.

Goal: Draw a straight line separating healthy patients from those with pneumonia.

2. Cardiology Risk

Math: Diagonal Threshold

Context: Predicting cardiovascular events.

X: Systolic BP. Y: LDL Cholesterol.

Goal: Learn the linear threshold separating high-risk patients based on combined vitals.

3. Oncology Biopsy

Math: Circular Boundary

Context: Evaluating tumor cell samples.

X: Cell Radius. Y: Cell Irregularity.

Goal: Draw a circular non-linear boundary separating benign (center) from malignant (outer ring) cells.

4. ER Triage AI

Math: XOR Logic Pattern

Context: Prioritizing ER wait times.

X: Patient Age. Y: Reported Pain.

Goal: Learn complex XOR patterns where high pain in extremes of age triggers critical flags.

Task Goal

Class 0

Class 1

---

Neural Architecture

Epoch

0

Loss (MSE)

--

Learning Rate

0.001 0.01 0.03 (Default) 0.1 0.3

---

Decision Boundary

The AI Engine Explained

Multi-Layer Perceptron (MLP)

This simulator runs a custom-built Supervised Learning engine entirely in your browser. It consists of an Input Layer (receiving patient metrics), Hidden Layers (where complex pattern recognition and non-linear transformations happen), and an Output Layer (the final diagnosis/prediction).

Forward Propagation & Activation

Data moves forward through the network. At each connection, the value is multiplied by a mathematical Weight (visualized by the colored lines) and shifted by a Bias.

Before passing to the next neuron, the sum is squished by an Activation Function. We use Tanh for hidden layers to allow negative values, and Sigmoid for the output layer to guarantee a probability between 0 and 1.

Backpropagation (Learning)

After making a prediction, the engine calculates the Loss (Mean Squared Error)—a measure of how wrong it was. Using calculus (the chain rule), it sends this error signal backward through the network, tweaking every single weight and bias by a tiny amount (the Learning Rate) so it makes a better guess next time.