Machine Learning and Cardiology
Fetal Peak Detection - Preprocessing
In order to achieve results in fetal heartbeat extraction, a signal should be preprocessed so that it is in a more appropriate form for separation into components. ECG recordings taken on the abdomen contain a high degree of noise, which is a result of resonance, echo and abdominal sounds. In these recordings, the fetal signal can be very weak, and in order to extract this signal, all other noise must be canceled out. The following sections explain the methods of preprocessing that are used.
3.1 Data
The signals used for fetal ECG extraction, come from a international Internet database known as the Physionet Database. It is a some what complete database of ECG signals that range from healthy subjects to those with conditions. In order to use the signals, the Biosig Matlab library was used. This allows loading of the specific file type that Physionet uses and gives a few functions to help in processing. Some simple preprocessing has been done in order to adjust the signals into a more manageable form.
3.2 Missing Values
One stumbling block that occurs in each signal is that some values are registered as `not a number' (NAN), causing issues due to the nature of the mathematics used in later calculations. One of the first approaches used was that all these values were set to zero. This was done for most of the first runs but what was noted to not be the most desirable method. What was eventually done was that the two closest non-NaN values on either side of the missing value were taken, averaged and the new value set to this. This made a smoother signal that would contribute to a better analysis.
3.3 Centering
One of the smaller pieces of adjustment done to the signal is known as centering. This is mainly done to help further more advanced methods of determining the fetal heartbeat. Its need comes from the fact that most, if not all, signals are above the zero line where an easy signal analysis into their relative frequencies can be done. Centering is achieved by taking the mean of all signal values and subtracting it from each individual value, thus achieving zero as the new mean value.
3.4 Denoising
Noise is a problem as it causes distortions in the signal that hinder signal extraction. The thoracic signals are especially so due to echoes in the chest cavity. In order to get a clean maternal signal from them, so that it can be compared to the abdominal signals to find the fetus, removal of the noise is crucial for more accurate results. This was done through the use of the Daubechie-2 wavelet, chosen for its similarity to a normal human heartbeat. A soft threshold technique was used when decomposing the signal and then to rebuild it to get a smoother heartbeat.
The main idea is to denoise the thoraic channels, which easens the detection of the QRS complexes and therefore also the synchronization of the different channels (since they will be synchronized based on these locations). See further for more details about this.
3.5 Baseline Wandering
The technique of baseline wandering helps to remove the inconsistent baseline in the original recordings. This was done using a moving average, leaving out the maternal QRS complex. After that, the baseline was reconstructed. When this baseline was subtracted from the original signal, a signal without any wandering in the baseline was achieved.