gms | German Medical Science

Introduction: Any sleep laboratory can choose from a number of vendors of software to record and store signals measured during sleep. These polysomnograms (PSGs) require a variety of required biosignals. These signals may be measured using multiple derivations and are complemented by additional signals. These biosignals need to be manually annotated by experts to detect sleep stages, arousals, respiratory events and leg movements. Many vendors use their own proprietary format for storing PSG records.

Although many vendors support European Data Format (EDF) export, which supports annotations [1], annotations are often not included and most vendors use a different proprietary file format. In addition, the format of the annotations and channel names are not standardised. This makes it difficult to create a pipeline for automated data analysis using different datasets, vendors or PSG setups. Therefore, we have developed a tool that can read multiple proprietary formats, harmonize PSG recordings and export them to an single EDF+ format with annotations included, that can be read by available open source software.

Methods: Our Python toolbox SleepHarmonizer uses a configuration-based approach. The configuration is based on a human readable YAML file that can be easily modified and extended. Data is loaded using dataset loaders, which provide all available signals and a list of harmonized annotation labels for a record in a specific dataset or from a specific vendor. The record loaders are open source modules listed in the Python Package Index (PyPI) and can be included in the configuration. This makes it easy to extend the toolbox to other datasets and vendors. The resulting dataset consists of individual EDF+ files per record.

Results: Currently, the toolbox is able to harmonize data from three public datasets used in our research (Sleep Heart Health Study, Multi-Ethnic Study of Atherosclerosis, MrOs Sleep Study)[@zhang2018national] and two specific vendors used at our local university hospital (Philips Alice and Somnomedics DOMINO). For datasets, all PSG setup information is included and no further configuration is required. For vendor software, a list of source channels is required, which may include fallback channel names if there are different setups. The configuration can be used to filter datasets by specifying record IDs, minimum sample rate by type and metadata such as apnoea-hypopnoea-index. Based on a comprehensive record loader configuration, channels can be typed, selected, renamed and derived. To create a new custom dataset, the configuration includes a list of target channels and annotation names. Source code and documentation are available under the MIT licence, see https://gitlab.com/sleep-is-all-you-need/sleep-harmonizer.

Conclusion: Due to the lack of standards for storing PSG data, our SleepHarmonizer toolbox can only harmonize data and cannot normalize the data to a specific standard. It is simple to use and can easily be extended by adding new datasets and vendors. The toolbox provides an easy way to create a dataset to analyse data from a sleep lab and add the value of reproducing any results found using public datasets. It has been successfully used for research into the automation of sleep classification and event detection [2], [3], [4].

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.