Joint Exploitation of Hemodynamic and Electrocardiographic Signals by Hidden Markov Models for Heartbeat Detection

Abstract

The detection of heartbeats is a fundamental task in the diagnosis of cardiovascular diseases. In a clinical setting, heartbeats are commonly detected from the electrocardiographic (ECG) signal, but it can also be estimated from hemodynamic signals, such as arterial blood pressure (ABP), pulmonary arterial pressure (PAP) and central venous pressure (CVP). In this paper, we conceived three HMM-based detectors that jointly exploit the information from ECG and ABP ($$ D_{ECG \& ABP}$$ ), ECG and PAP ($$ D_{ECG \& PAP}$$ ), and ECG and CVP ($$ D_{ECG \& CVP}$$ ). The HMM-based detectors are based on the comparison of the difference of log-likelihoods of observation of two competing models that learned the dynamics of the presence and absence of a heartbeat, respectively. The detection performances of the bivariate centralized detectors were similar, but the detector that simultaneously considers ECG and ABP provide the best result (sensitivity = 98.91% and positive predictivity = 99.11%). These results agree when a univariate HMM-based detector only considers the ECG signal in the decision process. The centralized bivariate detection approach proposed in this paper make uses of redundant information in the decision process and is particularly useful when one of the signals is too noisy or missing.