html`

Bottleneck MLP — model architecture

Input

UCI Arrhythmia

280 features

→

Dense layer 1

Activation

128 units · ReLU

→

Dropout

p = 0.4

→

Dense layer 2

Bottleneck

64 units · ReLU

→

Dropout

p = 0.4

→

Output

Softmax

16 classes

Dense + ReLU Dropout Softmax output Adam · lr=0.001 · Categorical cross-entropy

Input layer280 ECG-derived features from the UCI Arrhythmia dataset (452 patient records, 12-lead ECG morphology).

Dense layer 1 — 128 unitsFirst representation learning stage. ReLU introduces non-linearity. Wide enough to capture diverse ECG patterns.

Dropout — p=0.440% of neurons zeroed per training step. Aggressive regularisation essential for datasets under 500 samples.

Dense layer 2 — 64 unitsInformation compression forces the network to learn compact, hierarchical ECG representations.

Dropout — p=0.4Second regularisation gate. Prevents the bottleneck from memorising training data.

Output — 16 classes (Softmax)One neuron per arrhythmia class. Outputs calibrated probabilities summing to 1.0 across all 16 diagnoses.

Hover over each layer to learn more ↑

Accuracy vs. recall trade-off

Baseline model

Clinical failure

0 % accuracy

Global accuracy — deceptively high

Normal sinus rhythm recall0.95

Anterior MI recall0.00

Supraventricular tachycardia recall0.00

RBBB recall0.00

Every critical minority class recall = 0.00. The model never once detected a heart attack.

Final optimised model

Clinical winner

0 % accuracy

Lower overall — but saves lives

Normal sinus rhythm recall0.78

Anterior MI recall1.00

Supraventricular tachycardia recall1.00

RBBB recall1.00

Perfect recall on three life-threatening conditions. Powered by: class_weight='balanced'.

Pull quote

A false positive sends a patient for further testing. A false negative sends a patient home with an undetected infarction. These costs are not symmetric — and your model must not treat them as if they are.

Nuria Saka Bisoko · Multi-Class Cardiac Arrhythmia Classification · UOC