Prof. Dr. Ursula Gather
PD Dr. Michael Imhoff
The object of this project is to provide statistical methods for the development of intelligent alarm and analysis systems for bedside monitoring of the status of the critically ill. The data for this project is provided by the automatic online acquisition of vital signs monitoring data, laboratory data, medications and therapeutic interventions with a clinical information system on the surgical intensive care unit of the Klinikum Dortmund. In the near future an annotated reference dataset of vital signs monitoring data at one-second intervals will be available from a joint clinical study with the University Hospital Regensburg. A clinical alarm and analysis system for real-time application in intensive care has to be able to identify patterns of change such as level shifts, trends and outliers in physiological time series and to assess therapy effects. Data based knowledge has to be verified against medical expertise, and statistical methods have to be adapted and robustified for clinical online monitoring application. The statistical methods to be developed are in the fields of multivariate time series analysis and multivariate process control. The work in progress concentrates on the following topics:
Results achieved so far incorporate procedures for univariate process control of physiological time series, methods for dimension reduction, and obtained knowledge about associations between hemodynamic variables. These achievements provide for a solid background for future developments focusing on multivariate process control. Such control procedures should evaluate extracted clinical information with respect to their relevance and allow for an early detection of critical conditions. The application of univariate and bivariate rules for pattern recognition on weakly correlated signals seems appropriate for process control. The overall goal is to develop real-time algorithms that provide information about possible changes of state in the sense of pure change-point-detection, as well as qualitative and quantitative assessments of the patient’s medical condition. In the long run it is planned to implement the new methods in monitoring systems for use in medical practice. In preparation for this, the developed methods for univariate pattern recognition have to be validated in comparison studies against patient data. Therefore, an annotated reference dataset of vital signs monitoring data is currently acquired at the University Hospital Regensburg. The specified challenges are addressed in cooperation with projects A1, A4 and A5.
The theoretical results of the different packages should – at the same time – be verified against patient data in order to detect errors at an early stage of research. Verifying results requires the development of software code for the methods under consideration in collaboration with project A5.
Publications of the project can be found here.
The robfilter package was developed in collaboration with Prof.Dr. Roland Fried.It includes implementations of robust filters and sigal extraction methods for the statistics software R.
The package can be downloaded from CRAN.
For Details regarding the R project, please visit the official R-homepage. For details on Windows binary packages on CRAN, pleas visit CRAN/bin/windows/contrib.