Non-invasive fetal ECG using constrained ICA

Sæderup, Rasmus Gundorff

4. term

Signal Processing and Computing, Master

2018

2018-06-15

147

Ikke‑invasive metoder til at estimere fosterets elektrokardiogram (FECG) ud fra EKG-optagelser på moderens mave er attraktive, fordi de er mindre komplicerede og sikrere end invasive metoder. I dette speciale bruges independent component analysis (ICA), en statistisk metode til at adskille blandede signaler, for at udtrække et morfologisk korrekt foster-EKG. Vi afleder en constrained ICA (cICA), hvor optimeringen begrænses, så de estimerede kilder skal korrelere med valgte referencesignaler. Referencerne spænder fra simple pulssignaler ved forventede QRS‑tidspunkter (QRS er den skarpe spids i hjerteslaget) til skabeloner, f.eks. afledt af maternelle EKG’er, og der afprøves kombinationer af maternelle og føtale referencer. Vi gennemfører et hyperparametersøg, hvor skridtlængder, korrelationsgrænser og referencekombinationer varieres. Med de bedste indstillinger testes cICA på et syntetisk datasæt og sammenlignes med andre udbredte metoder, herunder template‑subtraktion med PCA samt ICA‑varianter som FastICA og Infomax. Resultaterne viser, at cICA kan udtrække foster‑EKG præcist, når det sande foster‑EKG bruges som reference, men præsterer dårligere med andre referencer og er da på niveau med klassiske ICA‑metoder. Ingen af de testede algoritmer kunne genskabe morfologiske mål (QT‑interval og T/QRS‑forhold) fra reelle EKG‑blandinger. cICA konvergerer hurtigt, men til et lokalt minimum, hvilket skyldes problemets ikke‑konvekse natur. Konklusionen er, at cICA ikke leverer morfologisk korrekte foster‑EKG’er, medmindre de sande foster‑EKG’er stilles som reference.

Non-invasive estimation of the fetal electrocardiogram (FECG) from ECG recordings on the mother’s abdomen is attractive because it is simpler and safer than invasive monitoring. This thesis uses independent component analysis (ICA), a statistical method for separating mixed signals, to extract a morphologically accurate fetal ECG. We derive a constrained ICA (cICA) that steers the optimization so that the estimated sources must correlate with chosen reference signals. These references range from simple pulse trains at expected QRS times (the sharp spike in a heartbeat) to templates, for example derived from maternal ECGs, and combinations of maternal and fetal references are tested. A hyper-parameter search varies step sizes, correlation thresholds, and reference combinations. Using the best settings, cICA is evaluated on a synthetic dataset and compared with established approaches, including template subtraction with PCA and ICA variants such as FastICA and Infomax. Results show that cICA can extract the fetal ECG accurately when the true fetal ECG is provided as a reference, but performs less well with other references and is then comparable to classical ICA methods. None of the tested algorithms could recover morphological features (QT interval and T/QRS ratio) from real ECG mixtures. cICA converges quickly but to a local minimum, reflecting the non-convex nature of the problem. The conclusion is that cICA will not deliver morphologically accurate FECG unless the true fetal ECG is supplied as a reference.

[This abstract was generated with the help of AI]

foster ; EKG ; elektrokardiogram ; ECG ; hjerte ; morfologi ; Independent component analysis ; ICA ; statistisk signalbehandling ; signalbehandling ; maximum likelihood ; FastCIA ; Infomax ; Vectorcardiogram ; Ikke-invasiv ; Gaussianity ; Informationsteori ; Negentropi ; Kalman ; Kalman filter ; PCA ; principal component analysis ; Lagrangian ; Kurtosis ; Optimering ; Filtrering

Download
View record in AAU Student Projects