Skip to content

Collaborative labelling of events in physiological waveforms

License

Notifications You must be signed in to change notification settings

MIT-LCP/metaann

Repository files navigation

metaann

Screenshot

metaann is a system for reviewing and labelling events in physiological waveforms. This system was built and used from about 2013 to 2017, to annotate several large collections of waveform data at the MIT Laboratory for Computational Physiology.

This system was built to solve a very specific problem. We have a collection of long-duration, multi-channel waveform recordings (ECG, PPG, blood pressure, etc.) And for each of those recordings, we have a list of events that were detected by some automated classifier. We have thousands of these events, and we want to group the events into buckets (e.g. "ventricular arrhythmia" versus "motion artifact") for later study. We also wanted to have several people review each event independently, to be sure they agreed.

As such, we needed a process that would allow a person to quickly review each event. It needed to allow the reviewer to focus on their task, which meant that the UI needed to load and display data without any perceptible delay, require only one click to submit an answer, and save the answers to a central server automatically.

The metaann client is a heavily modified, stripped-down version of WAVE. It uses GTK+ 2.24 (because that made it easy to port WAVE's Xlib display code to work natively on Windows and Mac OS.) Naturally, it uses WFDB, and libcurl.

The metaann server is implemented using CGI.pm. It relies on the PhysioNetWorks user authentication system. If you want to deploy it yourself, you'll have to provide your own authentication system and modify the scripts to work with it. Or write your own server; the HTTP API is not terribly complicated.

One significant issue we discovered - obvious in retrospect - is that not everyone works at the same speed. Some reviewers are highly productive and motivated, and can routinely review hundreds of events per week, while other reviewers will only review a handful of events now and again. If you have a specific target to achieve (e.g., maximizing the number of events that have at least two independent reviews), you can't simply divide up the database evenly among all the reviewers. The metaann system has no mechanism for addressing this situation, and required active involvement from the administrator (manually re-partitioning and re-shuffling the record assignment lists) in order to make effective use of everyone's time.

About

Collaborative labelling of events in physiological waveforms

Resources

License

Stars

Watchers

Forks

Packages

No packages published