Heart Risk Detector Neural Network

A feed-forward neural network implemented from scratch using NumPy to predict heart disease risk from tabular clinical data.

Features

• NumPy-only forward/backward propagation
• ReLU hidden layers, sigmoid output
• Mini-batch training with cross-entropy loss
• L2 weight decay and dropout
• Early stopping with validation monitoring
• Robust preprocessing: randomized train/val/test split, train-fitted normalization and one-hot alignment

How it works (Backpropagation and training details)

• Architecture: Input → Dense(16, ReLU) → Dropout → Dense(8, ReLU) → Dropout → Dense(1, Sigmoid)
• Forward pass: Matrix multiplies with biases (Z = A_prev @ W + b), ReLU in hidden layers, sigmoid for output probabilities.
• Loss: Binary cross-entropy with optional L2 penalty: BCE + (λ/2m) * (||W1||² + ||W2||² + ||W3||²).
• Backpropagation:
  • Output layer: dZ3 = A3 - y, dW3 = A2ᵀ @ dZ3 / m, db3 = sum(dZ3)/m.
  • Hidden layers: propagate gradients through ReLU using its derivative; if dropout is active, mask/scaled gradients are applied consistently to the corresponding layers.
• Optimizer: Vanilla SGD with learning rate lr (no momentum/Adam for clarity).
• Initialization: He/Xavier-style scaling (np.sqrt(2/fan_in)) for stable gradients.
• Mini-batching: Shuffles each epoch, trains on batches for efficiency and noise injection.
• Regularization: L2 weight decay on weights, inverted-dropout during training; dropout disabled during evaluation.
• Early stopping: Tracks validation loss; restores best weights after patience epochs without improvement.
• Prediction: Thresholds sigmoid probabilities at 0.5 to produce class labels.

Project structure

Heart_Risk_Dectector_NN/
  data/
    heart_data.csv
  nn/
    activations.py
    layers.py
    model.py
    __init__.py
  main.py
  README.md

Requirements

• Python 3.9+
• numpy
• pandas

Install with:

pip install numpy pandas

Dataset

Place heart_data.csv under data/. The script expects the target column named target, numeric columns ['age','trestbps','chol','thalach','oldpeak'], and categorical columns ['sex','cp','fbs','restecg','exang','slope','ca','thal'].

How to run

From the project root:

python main.py

This will:

• Randomly split data into 70% train, 15% validation, 15% test (seed=42)
• Normalize numeric features using train stats only and align one-hot columns
• Train HeartRiskNN with L2, dropout, and early stopping
• Print train/val/test accuracies and sample predictions

Key model settings (in main.py)

model = HeartRiskNN(
    input_size=X_train.shape[1],
    hidden_sizes=[16,8],
    lr=0.01,
    l2_lambda=5e-3,
    dropout_rate=0.3,
    seed=42,
)

model.train(
    X_train,
    y_train,
    epochs=600,
    batch_size=32,
    X_val=X_val,
    y_val=y_val,
    patience=50,
    verbose_every=100,
)

Overfitting fixes applied

• Randomized train/val/test split with fixed seed
• Train-set-only normalization and one-hot alignment
• L2 regularization and dropout on hidden layers
• Early stopping on validation loss
• Reduced capacity (hidden sizes [16, 8])

Example results

Example run (will vary by split/seed):

Training model...
Epoch 0: Train Loss=0.72, Acc=0.50 | Val Loss=0.70, Acc=0.52
...
Training Accuracy: ~0.86
Validation Accuracy: ~0.85
Test Accuracy: ~0.89