Fig 1.
Global architecture of the MFPD.
In the first two steps, data coming from ECG are filtered then converted into features Ci(t). and
are parametric probability models of feature i, representing valid and invalid detections.
Fig 2.
Specification of signal pre-processing and feature extraction (steps 1 and 2 in Fig 1).
For the application of robust QRS detection, a set of 3 features is extracted.
Fig 3.
Illustration of the multi-feature complementarity in fusion decision.
a) Raw ECG segment from record MIT-101 (SNR = 6 dB with electrode noise) b), c), d) represents the feature a, s, c respectively. The × and ⚬ indicate invalidated and validated QRS respectively by the fusion.
Fig 4.
Estimated distributions (dashed line) vs normalized histograms (vertical boxes) for the three features and for both and
.
Record MIT-101 is used with added baseline noise (SNR = 12dB).
Fig 5.
KLD variations for successive QRS candidates in 60 s from record MIT-231 with baseline noise and -6dB SNR.
Upper panel: raw ECG signal with annotations (marked by *), MFPD fusion results (+), single feature (SFPD) results (△, ⊲ and ⊳), and results of the reference methods (⚬ for UNSW, ◊ for SCD and □ for WBD). Lower panel: KLD evolution for three features in the same period.
Table 1.
Performances of the whole record MIT-231 of Fig 5.
Table 2.
Performance comparison on the benchmark MIT-BIH arrhythmia database.
Fig 6.
Se and +P (in %) of MFPD, UNSW, SCD and WBD on the benchmark noisy database, with different noise types and SNR levels.
Table 3.
Performance comparison on the benchmark noisy database.
Table 4.
MFPD detection delay (mean and std in ms).