Gaj Vidmar

Statistical Analysis of Uterine Electromyographic Data from Pregnant Sheep: Some Results, Problems, Possibilities and Suggestions

Abstract:
Based on the two years of my involvement as statistician, I will present an overview of results, problems, possibilities and suggestions regarding the analysis of data from over ten years of research of Prof. Pajntar's group on uterine smooth-muscle activity in pregnant sheep as a model for humans.

In that research, electromiography (EMG) was performed at the horn and the cervix of the uterus in more than thirty sheep, some pregnant and thus included in the experiment twice or thrice. After removal of unsuccessful measurements and digital filtering, the signals were processed to obtain root-mean-square (RMS) and median frequency (MF) over one-minute periods as the data for further analyses.

The research was set up to address a very broad array of issues: normal course of uterine smooth-muscle EMG activity with approaching labour, effects of mild electric stimulation on this activity, effects of labour accelerating or decelerating medication, and course of uterine smooth-muscle EMG activity shortly before, during and shortly after labour. This reflected the scarcity of previous research in that field, as well as the desire of the researchers to maximise the "data yield" with the given means. Hence, the design of the study was a mixture of cross-sectional observation, single-subject research with intervention and designed experiments, all with exploratory as well as confirmatory ambitions. Unfortunately, no statistician was involved in the design phase, so it was unavoidable for the study to show some weaknesses given the complexity of the issues.

Furthermore, the gathered data are extremely difficult to analyse for a number of mutually related reasons:

Nevertheless, we had some success in tackling some of the research issues. Firstly, pixelisation-based visualisation enabled us to produce a graphical summary of data quality and chronology of the entire research project. Next, robust descriptive graphics (boxplots of raw and aggregated data, local regression smoothing) gave some insight into the trends of normal course of uterine smooth-muscle EMG activity over the course of pregnancy. Linear modelling of transformed data (with mixed-model ANOVAs) enabled us to make limited inference regarding the effects of stimulation. In one of the experiments with medication, simple ANOVA with accompanying plots sufficed for confirming the ability of EMG to reflect the labour accelerating and decelerating drug effects, but in general, basic statistical methods are completely inadequate for our uterine EMG data.

It should also be stressed that the planned analyses with the final version of the data are not yet finished. Namely, until August 2005, the standard practice of the research group had been to use the bandwidth filter of 0.01 Hz to 4 Hz, but than it was narrowed to the bandwidth between 0.3 Hz and 3 Hz in order to improve the signal-to-noise ratio. In the same period, the final data screening for artefacts was completed.

For the future, we are putting hopes in interactive visualisation (with established information visualisation software packages for large datasets), which we already started, and in advanced longitudinal data modelling. But the essential step forward in importance of the results has to come from improved quality and usability of the input data for statistical modelling, and that has to come from improved signal processing.

In this respect, we will derive new data from the simple yet efficient approach of burst counting. I believe that the existing measures (one-minute RMS and MF) should also be complemented with peak frequency (mode estimated from a kernel-smoothed power spectrum) or a more informative power spectrum characterisation (perhaps based on mixture modelling, or clustering of spectra into a limited number of types, or on relative density of frequency bands specified on substantial ground). Further insight might be gained from measures derived via non-linear modelling (Lyapunov exponent, correlation dimension).

Another challenge for the future could be to automate the process of selecting useful measurements by identifying artefacts in the EMG recordings. All the data analysis efforts mentioned so far should, of course, in the final instance be directed towards reliable prediction of preterm labour, and perhaps even its prevention. However, even the basic data-analytic steps in this process are far from simple, and require notable dedicated time and resources, which have so far not been available.

Contact:
Gaj Vidmar M.Sc.
University of Ljubljana, Faculty of Medicine
Institute of Biomedical Informatics
Vrazov trg 2, SI-1104 Ljubljana, Slovenia
e-mail: gaj.vidmar@mf.uni-lj.si

Copyright 2007 You.com. All Rights Reserved.
Dreamweaver Templates Resources