Motivation

While PhysioNet  is a large database for standard clinical vital sign measurements [1], such a database does not exist for unobtrusively measured signals. This hinders progress in the vital area of signal processing for unobtrusive medical monitoring as not everybody owns the specific measurement systems to acquire such signals. Furthermore, if no common database exists, a comparison between different signal processing approaches is not possible. This gap will be closed by our UnoViS database. It contains different measurements in various scenarios ranging from a clinical study to measurements obtained while driving a car. Currently, 145 records with a total of 16.2 hours of measurement data is available, which is provided as MATLAB files or in the PhysioNet  WFDB file format. In its initial state, only capacitive ECG and unobtrusive PPG signals are, together with a reference ECG, included. The dataset from the clinical study contains clinical annotations. Additionally, supplementary functions are provided, which simplify the usage of our database and thus the development and evaluation of new algorithms. The development of urgently needed methods for very robust parameter extraction or robust signal fusion in view of frequent severe motion artifacts in unobtrusive monitoring is now possible with our database.

An example for the severe motion artifacts and the opportunities of a multichannel unobtrusive monitoring is shown in the followin figure. While the channel cecg₁ has a very low amplitude of the R-peaks even after recovering from a severe artifact, channel cecg₂ shows larger amplitudes of the R-peaks. Hence, it would be beneficial to use this channel for parameter extraction. Furthermore, the optical channels opt₁  and opt₃ recover even faster and would thus allow to increase the time in which a reliable heart rate could be estimated. This example clearly shows the opportunities of multichannel unobtrusive monitoring. Our database shall stimulate the research of very robust signal processing algorithms like peak detection, heart rate estimation or sensor fusion, in view of severe and frequent motion artifacts , as these methods are a key aspect for the success of unobtrusive, ubiquitous medical monitoring.

Terms of use

The database is completely free and can be used for any purpose. However, we kindly ask to cite this publication if the database and / or its supplementary functions are used:
T. Wartzek, M. Czaplik, C. Hoog Antink, B. Eilebrecht, R. Walocha and S. Leonhardt: "UnoViS: The MedIT Public Unobtrusive Vital Sign Database", Health Information Science and Systems. 2015;3:2. https://link.springer.com/article/10.1186/s13755-015-0010-1

Content

Currently, 145 records with a total of 16.2 hours of measurement data is available, which is provided as MATLAB files or in the PhysioNet WFDB file format. Additionally, supplementary functions are provided, which simplify the usage of our database and thus the development and evaluation of new algorithms. The data is available as a MATLAB mat-file or in several files per record in accordance to the PhysioNet WFDB (WaveForm DataBase) file format.

The origin of the data is manifold and currently consists of three application scenarios: measurements of a clinical study [2], measurements while the subject is driving a car [3], [4] and measurements while the subject is lying in bed [5]. Additionally, two records show the maximum measurement quality of our latest system if conditions are optimal. Since the measurements were acquired over a long time period of several years, slightly different (improved) measurement systems were used. In all scenarios, the monitored subjects wore their normal clothes.

Structure

The database consists of records which represent one measurement each. That means, a record contains, for example, one measurement of one patient in case of the clinical study, or one driver driving in a specified scenario in case of the driving tests and so on. It should be pointed out that several records may exist for the same subject in the driving or lying in a bed scenario. However, these records are of course labeled with the same unique subject ID but different record IDs.

The structure of each record in case of the MATLAB file is given in Table I. The presented structure of one record is saved as a struct. Each record consists of several fields such as an unique id, the duration of the measurement, the measurement scenario measScenario, information about the subject and several channels containing the actual measurement data and annotations ann.

The field ann is an array of structs within each channel. It contains the class (e.g. peaks or rhythm), the subclass which may be event based (e.g. in case of the class peaks) or interval based (e.g. in case of the clinical study the clinicians analyzed an interval of 5 s), the source (e.g. manual by medical experts or automatic by the open source ECG detector OSEA [6]) and the location loc of the annotation in samples. The type of the annotation(s) depends on the annotation class and subclass and is further elucidated in Table II. This table also shows all currently available annotations. All datasets contain detected peaks from OSEA or by a medical expert. In case of the measurements from the clinical study, this dataset contains the following annotations for the cECG as well as for the reference ECG. The diagnosed rhythm and if extrasystoles extrasys or a bundle branch block bbb are present. If the two clinicians could not clearly identify a parameter or if their results differed, it is denoted with NaN or ’-2’. Furthermore, the heart rate and different time durations such as PQ, QRS and QT time are given. Here, the mean value of the two clinicians’ results and the relative difference are given. Again, if one parameter, e.g. the PQ time, could not be clearly estimated, it is denoted with NaN.