WavePrep

WavePrep is a tool for creating machine-learning-ready signal datasets from MIMIC waveform databases. The tool allows for configuring pre-processing in a flexible, fine-grained, and channel-specific manner, supporting common signal preprocessing steps, such as downsampling, data cleaning, cutting out long nan-sequences, imputation, windowing, and splitting a dataset into training and test data. These steps can be specified via a user-friendly JSON file, lowering the entry barrier to using MIMIC waveform data. An exemplary performance analysis on a server with 80 available workers showed that the tool takes between 33.10 and 50.53 milliseconds to process 1 hour of recording when processing 6 channels. The modular, open-source tool enables researchers to add functionality according to their needs. Future work will extend this tool with a detailed dataset quality report, helping researchers create high-quality signal datasets with minimal coding. We plan to incorporate other open wfdb databases and user feedback. To use WavePrep, the user has to provide two input files. Firstly, a CSV file specifying the selected records, along with the time span for each record that should be included in the resulting dataset. Secondly, a JSON file specifying the channels to extract from the record, the channel-specific pre-processing steps, the windowing, and the split into train, test, and validation sets. Based on these two input files, WavePrep allows the user to create their custom machine-learning-ready dataset.