Application of supervised and unsupervised algorithms for automatic detection and classification of heart arrhythmias

[V. Atanasoski]

V, 86 str.

Biomedicina - Biomedicinsko inženjerstvo i tehnologije / Biomedicine - Biomedical Engineering and Technologies

Ova disertacija predstavlja studiju o novim metodama za pouzdano automatsko određivanje srčanih aritmija: ventrikularnih ekstrasistola (VES), atrijalnog flatera (AFL) i atrijalne fibrilacije (AF), iz signala snimljenih mobilnim EKG uređajem, i o novim metodama za obradu i uklanjanje šuma sa EKG signala. U prvom delu disertacije predstavljena je metoda za uklanjanje EMG šuma. Predloženi algoritam koristi novi iterativni pristup (IRM) čija je polazna hipoteza da se nakon uklanjanja dominantnih komponenti EKG signala (pre svega QRS kompleksa i T talasa) šum može jasno izdvojiti. Za potrebe validacije pomenutog algoritma, osmišljena je nova metoda koja simultano registruje EKG signale sa i bez prisustva EMG smetnji, te je kreirana baza podataka SimEMG koja je iskorišćena za potvrdu validnosti algoritma i objavljena kao javna baza. Signali iz ove baze su takođe iskorišćeni za analizu EMG šuma i optimizaciju pragova IRM algoritma. Kao značajan deo neophodan za detekciju aritmija, zasebno je predstavljena metoda za detekciju QRS kompleksa koja koristi više EKG kanala i dinamični prag za precizno identifikovanje QRS kompleksa, obezbeđujući pouzdanost pri različitim nivoima šuma koji se očekuju kod mobilnih EKG uređaja. Performanse ovog modela su pokazane na više javnih i privatnih baza, pokazujući veću osetljivost i specifičnost u poređenju sa konvencionalnim metodama. Za detekciju VES-ova razvijen je algoritam zasnovan na integralu QRS kompleksa u 3D prostoru, koristeći nenadgledanu DBSCAN metodu za klasterizaciju, dok su za otkrivanje AF-a i AFL-a razvijene tehnike dubokog učenja zasnovane na analizi RR intervala. Kvalitet ovih metoda je potvrđen na privatnoj bazi podataka i pokazana je efikasnost u razlikovanju normalnih i patoloških ritmova u EKG snimcima.

This dissertation presents a comprehensive study on methods for reliable automatic determination of cardiac arrhythmias: premature ventricular contractions (PVC), atrial flutter (AFL), and atrial fibrillation (AF) from signals recorded by a mobile ECG device, as well as new methods for signal processing and noise removal from ECG signals. The first part of the dissertation presents a method for removing EMG noise. The proposed algorithm uses an iterative regenerative method (IRM). The hypothesis behind the method is that after removing the dominant components of the ECG signal (primarily the QRS complex and T waves), the EMG noise can be easily separated. For the validation of the mentioned algorithm, a new method was devised that simultaneously records ECG signals with and without the presence of EMG noise. Additionally, the SimEMG database was created and used for the validation of the IRM algorithm and is publicaly available. The signals from this database were also used for the analysis of EMG noise and the optimization of IRM algorithm thresholds. As a significant part necessary for the detection of arrhythmias, a method for detecting the QRS complex using multiple ECG leads and a dynamic threshold to accurately identify the QRS complex is presented, ensuring reliability at various noise levels expected in mobile ECG devices. The performance of this model has been demonstrated on multiple public and private databases, showing higher sensitivity and specificity compared to conventional methods. For the detection of PVCs, an algorithm based on the integral of the QRS complex in 3D space was developed, using the unsupervised DBSCAN method for clustering, while for the detection of AF and AFL, deep learning techniques based on RR interval analysis were used. The quality of this method has been confirmed on a private database and has shown efficiency in distinguishing normal and pathological rhythms in ECG recordings.

elektrokardiogram, smetnje, elektromiogram, obrada signala, ventrikularne ekstrasistole, DBSCAN, atrijalna fibrilacija, atrijalni flater, duboko učenje

electrocardiogram, noise, electromyogram, signal processing, premature ventricular contractions, DBSCAN, atrial fibrillation, atrial flutter, deep learning

621.38:616.12(043.3)

Tekst