Given a biomedical signal such as an ECG, EEG, continuous blood pressure, pulse oximeter (PPG), ballistocardiogram or any other biomedical time series, researchers (or at least me) want to jump right to the feature extraction (heart rate estimation, detection of fiducial points etc.)
However, when working with real data, one quickly realized such signals are full of artifacts, e.g. due to movements, sensor and transducer failure to name a few.
What algorithms or preprocessing steps do you usually use to detect (and probably discard) such segments?
My pipeline usually goes as follows:
1. Physiological sanity checks, e.g. check that blood pressure reading above 30 mmHg and below 300 mmHg and that the pulse pressure is not too large. Make sure that between consecutive beats the statistics do not change too much etc.
2. Make sure there is some form of self-similarity in the signal (autocorrelation function) and that the spectral content is where we expect it (heart rate).
Any further hints and suggestions?
Thanks a lot!
Dear Federico,
You bring up a very relevant topic. In fact it is crucial for the whole field. In addition to you approach (which I like as highly practical): In case one has simultaneous recordings of several parameters, additional sanity checks can be done by looking e.g. at the modulation from respiration on signals as ECG-RR interval, PPG amplitude (although this effect strongly varies between body sites), Bio-Impedance (also varies with measurement location), etc.
A problem with multi-parameter recordings frequently is that each parameter has its own filtering chain, with individual differences in time delay. In the older days a millisecond more or less did not matter too much, but nowadays we start deriving more and more by looking at connections between various parameters in the time domain. But data recorded "simultaneous" mostly is NOT really simultaneous... This is one of the reasons that - to my opinion - it is of great additional value to measure a multitude of physiological parameters with truly synchronous time stamps and known filter chain delays for each parameter. And - if automatic adaptive filtering is applied - the associated delay characteristics are co-recorded.
For years I am an IEC member and have been writing standards for medical devices, but the topic you arise here Federico, indeed could use some more attention...
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Meta-Analysis