Classificação automática do registro do ECG de 12 derivações em aceitável ou inaceitável para laudo médico em sistemas reais de telecardiologia : redução de custos e de riscos no processo de diagnóstico cardíaco

Abstract

The automatic classification of the 12-lead Electrocardiogram (ECG) recording as acceptable or as unacceptable for a medical report is fundamental to minimize costs and risks related to the cardiac diagnostic process. Most algorithms for such classification include non-intuitive parameters and were developed based on data that are not representative of the real clinical scenario: underrepresentation of pathological ECG recordings and overrepresentation of unacceptable recordings. Therefore, new algorithms were introduced, which were developed and validated on data from a real telecardiology system - Telehealth Network of Minas Gerais (TNMG): Noise Automatic Classification Algorithm (NACA) to assess the quality of 12-lead ECG recordings; Limb Electrode Interchange Detection Algorithm (EIDALIMB) and Precordial Electrode Interchange Detection Algorithm (EIDAPREC) to detect limb and precordial electrode interchanges, respectively. These algorithms were developed based on clinical knowledge of the electrocardiographic tracing, using signal processing techniques of low computational complexity and using rules of physical or physiological meaning that are understandable and may be changed/adapted. The proposed algorithms were compared with other relevant ones in the literature using five metrics: Sensitivity (Se), Specificity (Sp), Positive Predictive Value (PPV), F2 and cost reduction resulting from the use of these algorithms in a real clinical system. Then, NACA was compared to the Quality Measurement Algorithm (QMA), winner of the Computing in Cardiology Challenge 2011, using data provided by TNMG (TestTNMG) and another publicly available, ChallengeCinC. On the other hand, EIDALIMB and EIDAPREC were compared to the Decision Rules Algorithm (DRA), which is currently the most renowned in the literature, using another dataset provided by TNMG (TestTNMGINV). TestTNMG and TestTNMGINV datasets consist of 34,310 and 23,235 ECG recordings, respectively (1% unacceptable and 50% pathological), while ChallengeCinC consists of 1,000 ECG recordings (23% unacceptable, higher than the actual clinical scenario). NACA and QMA achieved similar performance in ChallengeCinC, while only NACA achieved satisfactory performance in TestTNMG (Se=.89; Sp=.99; PPV=.59; F2=.76 and cost reduction 2.3 ±1.8%). On the other hand, unlike DRA, EIDALIMB and EIDAPREC achieved satisfactory performance in TestTNMGINV (Se≥.88; Sp≥.98; PPV≥.32; F2≥.63 and cost reduction 2.8± 2.2%). Therefore, it is noted that algorithms to be used in a real clinical scenario must be developed based on representative data of the clinical reality. Additionally, the implementation of NACA, EIDALIMB and EIDAPREC in any telecardiology service may result in evident health and financial benefits for the patients and the healthcare system.