Method Jimenez et al. (2002)

From Live review of baseline FHR methods
Jump to navigation Jump to search

Source

Jimenez, L., Gonzalez, R., Gaitan, M., Carrasco, S., Vargas, C. - Computerized algorithm for baseline estimation of fetal heart rate. - Computers in Cardiology pp. 477--480,2002

Method

The method is decribed to find the baseline on 5 minute long recordings. The FHR signal is sampled at one sample per beat. A first smoothing is realised by a moving-average of the FHR with a Hanning window of 27 points.

Then, the method extract same stable periods which will be baseline candidats. A baseline candidat is a period of at least 15s on which the derivative (in absolute value) do not exceed 1.0 bpm/s. Then, the average of the baseline candidat is calculated over the 5 min signal. The candidats which exceed the average + or - 10bpm are rejected.

A cubic spline interpolation between final candidats is then realised as well as linear extrapolation at both ends of the time series when necessary. At the end, a final baseline is obtained by applying a third order zero-phase low-pass filter with a cut aff frequency of 0.033Hz.

The accelerations (resp. decelerations) are defined as periods where the difference between the baseline and the original FHR indicate a segment of successive values with the same sign that reaches a peak (resp. a nadir) of at least 15bpm in less than 30s, and has a total duration of at least 15s.


Source code

The method was described to obtain a baseline on 5 min signal. In order to apply it on longer signal with correct results and for simplification some small adaptations have to be realised : - For computing simplification, the signal is sampled at 2Hz instead of 1 sample per beat which had no particular sense. - The process is realised on 5min sliding window to determine baseline periods and the cubic spline interpolation and the butterworth filtering is realised on the entire signal. - We are not completely sure on how is done the comparison between baseline segment and the average to eliminate candidats. - On some windows where two stable periods with important difference of FHR level, the global average of the window is then far from the one of both stable periods. Consequently, all baseline candidat are rejected and the is no baseline periods on the entire window. This leads to huge period which are interpolated by spline and the interpolated signal is often aberrent (e.g. can be 200bpm over the maximum of the FHR). We discover that taking the median of the baseline candidats is much more efficient than taking average and it gives results closer to the one described on the paper.

Here is the specific source code of Jimenez computation for baseline, accelerations and decelerations. The required files do visualise it are available on our FHR analysis librarary.


Discussion

The idea to reject unstable period on the estimation of the baseline is insterresting and the method is efficient for removing short accident for baseline estimation. Unfortunately, this only work on short duration accident with clear begining.

Jimenez worked only on 5min signal which is unsufficient to detect prolonged deceleration. In addition, Jimenez validates accident only if the nadir (deceleration) or peak(acceleration) is reach before 30s which is the definition of variable deceleration but it is not the case of all deceleration/acceleration.

In addition, the cubic spline interpollation is a bit unstable and linear interpollation would probably often give better results. As example, if a deceleration start and finish slowly, the begining/end of delecelation will be considered as stable periods and will be used as support for interpollation. The estimated baseline will then follow the decrease insteal of going straight.

File:Jimenezspline.png
Example of spline interpolation result on a deceleration with slow begining/end