Scientific Background

Why ECG quality is a modelling problem

An ECG analysis system can return a confident result even when motion artefact, poor electrode contact, baseline drift, saturation, or signal loss has removed the morphology needed for interpretation. Signal-quality assessment therefore acts as an input-validity check before heart-rate, rhythm, or morphology analysis.

This project follows the classical idea that no single quality mechanism is sufficient. Instead, multiple imperfect indicators are combined into a quality decision. The reproduced study is Clifford et al. 1.

Seven SQI families

For a 12-lead record, each SQI is computed per lead, producing an 84-dimensional representation.

SQI	Mechanism represented	Typical failure signal
bSQI	Agreement between independent QRS detectors	Missed or inconsistent beats
iSQI	Inter-lead agreement	Lead-specific corruption
kSQI	Kurtosis	Abnormal waveform-shape distribution
sSQI	Skewness	Asymmetric shape distortion
pSQI	QRS-band spectral concentration	Spectral contamination
fSQI	Flat-line occupancy	Signal loss or saturation
basSQI	Low-frequency power	Baseline wander

The fusion view can be written as

\[ z = \phi_{\mathrm{SQI}}(x), \qquad P(Y=\mathrm{poor}\mid x) = F(z_1,\ldots,z_K). \]

The individual features remain interpretable, while the classifier learns how their failure modes overlap.

Model families

RBF-SVM. A nonlinear margin classifier used for individual SQIs and their predefined combinations. It is the closest classical comparison with the paper.

LM-MLP. A small multilayer perceptron trained with a Levenberg-Marquardt-style procedure. It tests whether flexible feature fusion changes the classical conclusion.

ResNet and Conformer. Waveform models preserve local temporal evidence. The matched ResNet is an architecture control: on BUT QDB it remained comparable with the Conformer, preventing a universal architecture-specific claim.

Datasets and evaluation

Dataset	Role	Labels used
PhysioNet/CinC 2011 Set-A	Public 12-lead reproduction and extension	Acceptable / unacceptable
MIT-BIH Noise Stress Test Database	Electrode-motion and muscle-artefact controls	Noise recordings
PTB-XL	Clean carrier signals for train-only proposal construction	Public waveform data
BUT QDB	Native single-lead graded evaluation	Good / Medium / Bad

All augmentation is restricted to training. Validation and test partitions remain native. Accuracy is reported with class recall, balanced accuracy, ROC-AUC, PR-AUC, and macro-F1 as appropriate; these complementary metrics separate ranking quality from a selected operating point.

Validity principle

Balancing the class prior, \(P(Y)\), does not guarantee recovery of the poor signal distribution, \(P(X\mid Y=\mathrm{poor})\). The project therefore audits whether generated and native poor ECGs can be distinguished and whether models trained on one source transfer to the other.