Model Architecture

Network Design

The neural network was designed to balance capacity and generalization for a 13-feature tabular dataset.

Layer	Neurons	Activation
Input	13	—
Hidden 1	64	ReLU
Hidden 2	32	ReLU
Hidden 3	16	ReLU
Output	1	Sigmoid

Loss function: Binary Cross-Entropy
Optimizer: Adam
Regularization: Dropout after each hidden layer

Why 3 Hidden Layers?

The choice is grounded in the bias-variance trade-off:

Too few layers → underfitting (high bias), the model cannot capture meaningful patterns.
Too many layers → overfitting (high variance), the model memorizes training data.

Three layers sit at a reasonable middle ground for a dataset of 1,025 samples with 13 features. Alternative architectures (1 hidden layer and 5 hidden layers) were also tested to validate this choice.

Why Sigmoid on the Output?

The output layer uses a single neuron with a Sigmoid activation, which squashes the output to the range [0, 1]. This directly represents the probability of heart disease being present, making it natural for binary classification with Binary Cross-Entropy loss.

Regularization with Dropout

With only 1,025 training samples, overfitting is a real risk. Dropout randomly deactivates a fraction of neurons during each training step, preventing the network from co-adapting and forcing it to learn more robust representations.