foetal phonocardiography (PCG) signal

The origin of the PCG signal

PCG signals were recorded from pregnant women, during the last month of their singleton physiological pregnancies (34-40 weeks of gestation). In this paper[1], they used a portable phonocardiograph, ‘Fetaphon Home’, equipped with a microphone for FHS recording along with a traditional pressure probe for uterine contractions (US) recording. The sampling frequency of the phonocardiographic device is 333Hz.

FHS signals

In FHS signals first and second foetal heart sounds (S1 and S2, respectively), shaped as bursts, can be recognized. S1 is considered a good time indicator of heart beat, because of both its high energy with respect to the other portions of the FHS, and less morphologic variability .

Fig.1. Example of FHS signal (PCG recording #12) in wich visually recognized S1 and S2 sounds are marked as 1°S and 2°S. The zoom shows the S1 template chosen in this case.

An algorithm for FHR estimation in this paper[1]

This algorithm was basically designed on S1 enhancement and detection. A logic block, based on mathematical rules mainly concerning mean heart beat characteristics, described in detail in the follwing B4(“Logic block”), chose S1 among serval ‘candidates’. Finally, a post-processing block determined the time position of S1. The flow chart in Fig.1 represents the mian steps of the signal processing.

Fig.2. Main blocks of this algorithm. Some blocks (B2 and B3) were implemented in different ways, as explained in detail in the text.

The main process of this algorithm

B1. Pre-processing

At this step, they over-sampled PCG signal (re-sampleing frequency equal to 1332Hz) in order to reduce errors due to low finite sampling rate.

B2. Filtering block

B2.1. Matched filtering

In matched filter design, the underlying assumption is that the time event has a known wave shape. There is an important difficulty in using matched filtering in FHS processing since the actual wave shape of the foetal hearbeat is unknown, and so its estimation is necessary. This problem was solved here, for each PCG resording, using a template of S1, manually selected by the operator in a portion of FHS signal with high SNR. An example of FHS signal and chosen template is reported in Fig 2.

B2.2. Set band-pass filtering.

In PCG recording, interferences are confined below 20Hz (mostly internal noise, for example maternal heart sounds and digestive sound) and above 70Hz (environmental noise). However, it is separable in time-domain since time relationship between S1 and S2 is known; typical time distance between S1 and S2 peaks is about 150ms.
in this paper[1], the filtering was designed by using a digital band-pass filter having a 3dB band equal to 34-35 Hz and centered at 44 Hz.

B2.2. Variable band-pass filtering.

they designed a digital band-pass filter with bandwidth equal to 20Hz (as for step B2.2) and central frequency chosen according to week of gestation of PCG recording (see Table 1).

B3. First heart sound enhancement block

This step was necessary because disturbances still remain in the filtered PCG signals. Hence, at this step, the algorithm had the main goal to discover the presence of bursts corresponding to S1 in FHS signals.

B3.1. Teager energy operator (TEO)

TEO is a nonlinear time operator, defined in both the continuous and discrete time-domain.
The following expression is used to describe the TEO in discrete time-domain:
$$E(n)=x^2(n)-x(n+1)\ast x(n-1)$$
Owing to the residual noise, the B3.1 block output was low-pass filtered, by using a low-pass filter with a cut-off frequency of 30Hz.

B3.2. Autocorrelation

Autocorrelation technique was chosen because it has the property of emphasizing periodic components in noisy signals while reducing non-periodic components (like movement artifacts, foetal movements, casual environmental noise).
They computed the autocorrelation envelope (B3.2 output) by means of cubic interpolation of all local maxima (see Fig.3 for an example).

B3.3. Autocorrelation+TEO

In this combination, the overall autocorrelation signal should increase SNR by suppressing the non- periodic noise in FHS; then, the TEO operator should further enhance heart sounds and slightly smooth the autocorrelation result.

B3.4. TEO+Autocorrelation

In order to reduce remaining noise, outputs of these last blocks were further low-pass filtered. Just for simplicity, we will call LP the output of B3 block.

B4. Logic block

This block had a crucial role for the developed algorithm. It had the aim to identify all (and only) the lobes of LP corresponding to the possible location of S1.
It is worth stressing that, due to noise, there was not a one-to-one correspondence between LP lobes and S1 and that some lobes could show multiple peaks(see Fig.4)

Lobes were analyzed with regard to their inter-distance consistency (position criterion) and amplitude regularity (amplitude criterion). Peaks with a position in a fixed time interval (based on the mean value of peak inter-distances computed respect to the previous 8 detected beats) were considered as candidate beats. Among candidate beats, those with amplitude greater than a fixed threshold (chosen with reference to the average of the maximum amplitudes of the previous 8 detected beats) were considered as probable heartbeats, because peaks satisfying both criteria most likely preserve continuity of the FHR series.
In the T time interval there could be more,one or no peak. To select probable heart beats, the algorithm marked only peaks higher than the amplitude threshold estimated on the previous 8 detected beats (50% of the mean of their maximum amplitudes, here called high threshold:HT). If just one candicate beat satisfying the amplitude criterion was present in the interval T, the algorithm recognized it as detected beat and utilized its characteristics to update T0, MEAN and amplitude threshold values and then repeated the cycle (up to signal end).
Otherwise, if more peaks higher than HT were present in T(probable beats), called Pi their position, the algorithm chose the peak that minimized the distance|T0-Pi-MEAN|; that is we considered it like the most probable peak because its position was the closest to the previous mean value.
If in T there was not any peak greater than HT, the algorithm repeated the search setting the amplitude threshold at 30% of mean of maximum amplitudes of the previous 8 beats,(here called low threshold:LT).
If, also with this condition, in T there was not any candidate beat satisfying the amplitude criterion, in order to avoid missed beats, the algorithm fixed a peak in the time position P0=T0+MEAN and associated to this peak the lowest fiducial degree .
At the end, the logic block outputs are the recognized S1 associated with an index showing satisfied criteria; it will be used by the post-processing block.

At the end, the logic block outputs are the recognized S1 associated with an index showing satisfied criteria; it will be used by the post-processing block.

B5. Post-processing block
This block had different task: to locate time marks for detected S1, to estimate FHR signal, to compute a reliability index and to identify and substitute outliers.

[1]: An algorithm for FHR estimation from foetal phonocardiographic signals(2010).