Hidden semi-Markov model-based heartbeat detection using multimodal data and signal quality indices

The automatic detection of heartbeats within physiological signals collected from patients connected to bedside monitors is an important task as it allows the detection of pathological conditions. Heartbeat detection is traditionally performed using the ECG. However, all bedside monitors are prone to missing data, yet it is rare for any system to incorporate data from other cardiac signals, such as the arterial blood pressure (ABP) or photoplethysmogram waveforms. This paper discusses the development of an automatic heartbeat detector using multimodal data from bedside monitors for the Physionet/Computing in Cardiology Challenge 2014. The presented algorithm employs an extended hidden Markov model to identify beat locations from multimodal data. The model was extended to include F1-score based signal quality indices in order to identify noisy periods. Wavelet transform features from both the ECG and ABP signals were added to derive the probability of a beat being present at a given location. The overall score of the algorithm for the third phase of the Physionet Challenge 2014 was 83.47%. The algorithm was also evaluated and compared to the top ranked entries [1] on a sample of 5150 synchronous ECG and ABP records from the MGH/MF database [2]. The overall score in this case was 92.7%.