Finding relevant cases in large databases of signals, time series, and clinical data

To make effective use of web archives such as PhysioNetpsilas multi-terabyte collections of ICU recordings, investigators need robust and flexible tools for finding data relevant to their studies. We have designed and implemented web-based visualization and search tools for this purpose. In addition to selecting recordings with desired demographic, clinical, and technical features, the search tool permits one to locate recordings containing events defined by characteristics of the recorded signals and time series, such as changes in parameters of user-specified amounts over user-specified time windows. By precomputing signal features, and by caching intermediate results, such searches can be made far more efficient than brute-force methods permit. At present, searching a collection of over 17,000 patient recordings requires about 4 seconds for an average query, thus allowing investigators to find useful subsets of very large data collections without major computing efforts.