Method Taylor et al. (2000)
Taylor, G. M., Mires, G. J., Abel, E. W., Tsantis, S., Farrell, T., Chien, P. F., Liu, Y. - The development and validation of an algorithm for real-time computerised fetal heart rate monitoring in labour. BJOG: an international journal of obstetrics and gynaecology 107(9):1130--1137, sep 2000
The method to determine the baseline is a mutlistep method based on the progressive trim principle. - Preprocessed signal : x - Set b0 a first baseline signal obtained by low pass filtering x with a third order Butterword forward-backward filter at cut off frequency 0.008 Hz. - for i from 1 to 3 -- compare bi-1 and x and consider periods with more than a thresold Ti bpm difference as accident period (either deceleration or acceleration). Ti is set to +-5bpm for the two first iteration and to +10bpm -5bpm for the third iteration . -- Set Bi as x but without accident periods which are linearly interpolated. -- get an intermediary baseline bi by filtering Bi with a low pass filtering with a third order Butterword forward-backward filter at cut off frequency 0.006 Hz.
The final basline is B3.
Once the baseline is determined - The baselevel is determie by the mean baseline on the segment. - Accident is drawn on the signal X which is obtained by low pass filtering with a 4th order Butterword forward-backward filter at cut off frequency 0.02 Hz. We suppose that the accident periods are defined to start when there is +10bpm or -5bpm difference between B3 and x and we suppose that those accidents are validated if they have an amplitude greater that 15bpm and a duration greater than 15s measured between B3 and X, but our only clue is that it is based on FIGO reccomendation. - The FHR variability is determine as twice the standard deviation of the FHR after removing the baseline on periods outside accelerations and decelerations.
Women 14 active labour and 16 undergoing labour. Duration 143 min to 1155 min with a median 484 min. -> random : 24 four CTG of 25min
Seven experienced reviewer (senior obsteric staff)
Criteria : - base level (correlation 0.93 between experts; computer vs expert : 0.91 to 0.98) - number of deceleration : (correlation 0.93 between expert; computer vs expert : 0.82 to 0.92) - number of late deceleration : (correlation 0.79 between expert; computer vs expert : 0.68 to 0.85) - number of acceleration (correlation 0.27 between expert; computer vs expert : 0.06 to 0.80) - baseline variability : >=5 or <5 : (kappa=0.27 between expert; computer vs expert : 0 to 0.34)
The idea of progressive trim is quiet efficient and we suppose the method work on most of signal. We guess the limitation can be that the begining and the end of deceleration/acceleration can be shifted which is not evaluated on the paper. This shifting can be due to a possible forget of describing how the begining/end of accident is calculated is determine (if they were based on derivated signal the shifting could be reduced) and to slight erroneous baseline estimation on the border of accidents.
In addition, this method can have difficulties on prolonged deceleration and when duration between deceleration is short.