Kevin McGill

    This is an appropriate time for a symposium on EMG Decomposition, because 2007 marks the 25th anniversary of the famous paper by LeFever and De Luca that really started the modern era of computerized EMG decomposition. There has been continual interest in decomposition since then. Many advances have been made, and decomposition has been increasingly used in scientific and clinical studies.

    And yet, if we look at the field as a whole, we see that even after 25 years EMG decomposition is not widely used outside of a dozen or so laboratories. There is little intercommunication between labs. There are no agreed-upon standards for exchanging data. There is no widely available software. And, there still remains in the wider scientific community a certain amount of skepticism about whether EMG decomposition is really as accurate and reliable as we would like to think that it is.

    So to try to remedy this situation, a group of us got together at last year's ISEK meeting—in the hallway between sessions—to inaugurate a Special Interest Group with the goal of trying to see if there aren't ways we can work together to advance the art and practice of EMG decomposition.

    I'm delighted that we've gathered together now for the first meeting of this group. The theme of the meeting is "Working Together." and I'd like to outline three areas in which I think that by working together we can make some important advancements.

Data Sharing

    The first way we can work together is in the area of data sharing. We all know that one of our responsibilities as investigators is to publish and disseminate our results. There is a growing awareness that this responsibility extends to sharing data as well. Fins and Gardner of the Cornell Neurodatabase Project make the case for data sharing very strongly: "Because science is a collective enterprise, producers of data are obliged to share their products with their peers for review and utilization for the advancement of knowledge."

    Data sharing offers several tangible benefits both to those who donate data as well as to those who access it. It provides the most definitive way for investigators to document their scientific findings. It provides a valuable resource to help guide the development and testing of new algorithms. It is also an excellent way to demonstrate the capabilities of decomposition, so that the skeptics can see for themselves.

    So what can we do? I'd like to suggest that we can work together to establish a useful, accessible, high-quality database of EMG signals.

    We have already made a start at this. Over the last year we have established a web-based forum—http://www.emglab.net—for issues related to EMG decomposition, and one of its most important features is a signal database. Currently the database has several small datasets illustrating the types of signals that are of interest in different labs. It also contains two more substantial datasets: a large set of clinical signals from Miki Nikolic, and a large set of simulated signals from Andrew Hamilton-Wright. It also contains datasets that were used for validations purposes in several papers. The site contains an online signal viewer, and many of the signals are annotated.

    If the data in the database is to be generally accessible, then the issue of data format arises. Different investigators collect EMG signals in different formats because of the particular hardware and software they use. We've decided to take the approach used by Physionet.org, an online database of physiological signals. We post the signals in their original formats and supply for each signal a simple text header file that contains information about the byte ordering, number of channels, sampling rate, gain, and so on. We also provide Matlab code for reading data files based on the information in the header.

    When it comes to the output of decomposition—the MUAP waveforms and lists of firing times, which we refer to collectively as "annotations"—a standard format is not currently available, but would be very beneficial for facilitating the exchange of information between labs. Ted Clancy has put together a draft annotation format that we think will be useful and versatile. We're now seeking input and feedback from the decomposition community, and we hope to have reached a consensus by our next Special Interest Group meeting next year.

    The other thing that is necessary to make the database useful is a set of reliable annotations. This is especially important if the signals are to be used for validating new algorithms. For real signals there is no way to be absolutely certain that annotations are correct. But we would to have at least a subset of the annotations carefully checked by individuals from more than one lab. Annotations checked in this way will be given a "blue ribbon" seal of approval, indicating the fact that they have been independently checked by more than one lab. Over the course of this year we will be asking people to participate in a trial "blue ribbon" panel.

Software

    The second area in which we can work together is software. There is currently no widely available software for EMG decomposition. Each lab generally uses the software that they themselves developed over the years. This makes it very hard for new investigators to start doing decomposition, or for even established investigators to take advantage of newly published algorithms. There have been a couple of commercial products, including Motor Unit Tools and Precision Decomposition, but they have not been widely successful because of the small market.

    (I am speaking here primarily of decomposition for physiological studies, not clinical studies. There is a potentially large market for clinical EMG decomposition, but the main problem here is to gain physician acceptance. Dan Stashuk and Miki Nikolic are working hard to accomplish this.)

    So what can we do? I'd like to suggest that we can work together to develop a common software infrastructure that will make it easer to implement and share new algorithms. The idea is to develop an open-source infrastructure and set of modules. Then anyone who is interested in a particular aspect of decomposition can make improvements to that specific module, and the rest of the community can benefit.

    We've started to work in this direction by taking the EMGlab decomposition program that was developed over the past ten years in Palo Alto, rewriting it to make it more universally useful, and making it available as open-source Matlab code. Some of its features include the ability to read both single- and multi-channel signals in a variety of formats and sampling rates, a convenient graphical user interface for viewing signals and manually editing annotations, a powerful algorithm for resolving superimpositions, and a basic automatic decomposition algorithm. The program can be used to decompose signals in its own right, and it has now been downloaded over 200 times.

    We hope that the program will provide a useful framework for future software and algorithm development. We've already had one major contribution—Joel Florestal has made his multi-channel decomposition algorithm available, and it has now been integrated into EMGlab. We look forward to more contributions in the future.

Quality Assurance

    The third way we can work together is in the area of quality assurance. Decomposition is a complicated process. It involves complicated algorithms and sometimes subjective decisions, and there are numerous points at which errors can arise. We do our best to get it right, but we all know that some results are more confident than in others. Unfortunately, this is seldom reflected in our papers. Therefore it should come as no surprise that some readers unfamiliar with decomposition may see out results as little more than wishful thinking.

    So what can we do? I'd like to suggest that we can all pay more careful attention to reporting the accuracy of our decomposition results. The way we often do things now is to perform a validation study involving a limited set of signals early in our careers and then refer to it forever after. But we really know that accuracy can vary from signal to signal and even from train to train.

    Most scientific measurements are shown with error bars to indicate the level of uncertainty in the measurement. I'd like for us to figure out how to provide error bars for our decomposition results. Then we could report, for example, that the results in a particular study were based only on trains for which there was a 95% level of confidence that at least 98% of the discharges were identified correctly to within ± 0.5 ms. I believe that the use of objective estimates of decomposition accuracy such as this will be an important step in documenting the scientific integrity of our work.

Conclusion

    In summary, I've tried to present several ways that we can work together to make the field of EMG decomposition more scientifically sound and more widely accepted. I hope that through this Special Interest Group we'll be able to make some of these things come about.