Artikel
A self-learning expert system and knowledgebase for neurosurgical patient management
Ein selbstlernendes Expertensystem für das Management neurochirurgischer Patienten
Suche in Medline nach
Autoren
Veröffentlicht: | 4. Mai 2005 |
---|
Gliederung
Text
Objective
In a retrospective study 2485 patients admitted to our clinical unit from January 2002 to November 2004 were analyzed. The aim of our study was to create a self-learning expert system and knowledge base, giving support for treatment planning, defining clinical pathways, analyzing risk factors and adverse events, optimizing DRG- (Diagnosis Related Groups) and procedural encoding.
Methods
A database was set up including data for planning and accompanying the in-patients stay in our hospital. Additional data was collected including history, diagnosis, factors of risk, first symptoms, neurological state, description of CT and MRI imaging, Karnofsky Performance Index, MacDonald Criteria, planning of treatments and surgery, treatment recommendations and adverse events. For patients, who underwent surgical procedures, all related data was added including adverse events. DRG and procedural encoding are obtained from the hospital’s patient record. A plain text semantic analysis in the database and in patient-related documents was performed extracting 4621 different tokens together with its related statistical information.
A set of system 2310 variables is defined, statistical analysis is performed for each variable and the statistical relations between all variables are determined. In the next step, a factor analysis is carried our,thereby reducing the number of significant variables to 732. For each of these variables the dependencies on all other variables is analyzed, calculating the covariance vector for the dependent variable and the covariance matrix of all independent variables and performing a linear “Multivariate Regression Analysis” using the Gauss Jordan algorithm, resulting in a vector of regression coefficients.
From these coefficient vectors, a matrix is built and stored in a knowedgebase containing the rules of transformation from a set of variables derived from a real patient to a set of 732 predicted values. During this global analysis, tables are generated for each of the variables showing its dependencies on other variables including correlation, covariance, and other descriptive statistic values in order of decreasing correlation.
Results
For new patients, the expert system performs analysis by extracting all available data from different databases and generating a set of 732 prediction values using mathematical transformation. Those variables are selected where the normalized difference between the prediction value and the actual value for the patient is beyond predefined thresholds. The results are presented as a text document in the descending order of the normalized difference between the predicted value calculated by the expert system and the statistical mean value. This patient-related analysis takes about 2 seconds on a standard 2000 MHz Pentium 4 Processor.
This study showed that it is possible to implement and to train an expert system for neurosurgery. Statistical data obtained during analysis and training is useful in global risk analysis and in the definition of clinical pathways by showing up factors of influence and dependencies. The expert system makes useful propositions for establishing optimized treatment plans proposing examinations, procedures and other treatments to be done, warning of planned treatments or procedures that must be avoided, giving alert for patients with increased risk, predicting time for surgical procedures, making propositions for correct DRG- and procedural encoding and avoiding inappropriate encoding.
Conclusions
This study shows that it is possible to describe medical knowledge and neurosurgical expertise by mathematical methods, at least to a certain degree.