Clinical decision support
| industrial collaborators: | E-Tech Limited |
| academic collaborators: | University of Durham |
| initiated : | 2003/04/20 |
| last updated: | 2009/08/27 |
 |
 |
Project summary: Clinical decision support
This is a summary of a Faraday Partnership research project on clinical decision support involving the University of Durham and E-Tech Ltd. The link at the bottom opens the summary as a two-page pdf document.
Overview
Within orthopædics, clinicians routinely take multiple measurements on patients during the course of their treatment, often repeating the same measurements before and after operations, and subsequently at periodic follow-up consultations. This data, combined with additional factors, comprises a high-dimensional data set with a mixture of data types and repeated observations over time. Statistical analysis of such data can be problematic.
Typically an orthopædic consultant would not have access to the statistical know-how needed to extract full value from the data. A tool-kit of automatic or semi-automatic methods would enable investigation of the data with minimal effort to reveal patterns, compare patient groups and predict outcome of future treatment for new patients. To support the development of a tool-kit for clinicians, this project has developed general statistical methods for the investigation and analysis of a generic orthopædic data set.
Technical summary
The focus of this project has been the analysis of clinical outcome data from total hip and knee joint replacement. However, a key purpose of this work was to tackle such data sets in generality. This necessitated the construction of a structural framework at an early stage, which was imposed on the types of data sets to be considered. This framework gave structure to the possible analyses one may seek to perform on these data sets, and enabled subsequent development of methodologies. Data sets to be analysed typically comprise between 500 and 2500 patients and in excess of 100 patient variables with between 10 and 20 variables observed repeatedly over time.
To support initial explorations of the data, and in order to gain basic insights into the nature of these data sets, various graphical methods were developed. These included the t-test plots, and standardised profile plots, shown in Figure 1 for the hip replacement data. These methods were considered both effective and intuitive in a clinical setting. The profile plots were especially well received as they give an immediate overview of the patient’s mean evolution over time on many different measurements. An initial comparison of multiple subgroups of patients was also made.
Figure 1: Profile plot of 6 variables of the hips data set split according to NHS status. The three points correspond to the standardised mean patient state for each group observed pre-operatively, and at 3 and 12 months post-operation.
The second strand of the project was the development of a novel method for selecting a subset of the variables that retains a high proportion of the total variation in the data. This is important due to the high dimensionality of the data and the replication of measurements over time. The approach taken has been demonstrated to be highly efficient and effective in comparison to other methods in the literature. The extension of the selection process for repeated measures and utilities attached to variables played a prominent role in the subsequent analyses.
The final problem to be tackled was that of the modelling of the data. Graphical modelling and chain graph models were chosen for their intuitive interpretation in terms of statements of conditional independence, and because they were readily able to accommodate the repeated measures structure of the data. An example of a final model for the knee replacement data is shown in Figure 2. The presence or absence of important predictive relationships can be identified easily by the clinician from this intuitive representation.
Figure 2: Chain graph for the knees data. Filled and unfilled circles correspond to discrete and continuous variables respectively. Arrows indicate dependence between pre-operative (top row) and post-operative (bottom row) measures.
Overall, the methods presented here have met with success, enabling the almost routine analysis of a generic orthopædic data set which conforms to the abstract structure. Combination of the various strands of research provides a method for exploratory analysis through dimension reduction, modelling and prediction. As a future development of this work, all of these methods could be combined with a simple data management system to form the basis of a statistical support tool-kit for clinical decision-making in orthopædics.
"This project has produced methods that are novel and provide an intuitive presentation of the outcome of orthopædic joint replacement for a clinician to understand without the need to comprehend the underlying complex analysis. This provides clinically useful methods to work with the multiple parameters that describe clinical outcome and retain the fine detail that is lost by simply adding the various terms together. The next step will be to build an outcomes analysis tool employing these methods and testing this in a clinical environment."
John Egan, E-Tech Ltd
References
- J A Cumming and D A Wooff (2006) Dimension reduction via principal variables. Computational Statistics and Data Analysis, in submission.
- J A Cumming and D A Wooff (2006) Standardised profile plots for multivariate repeated measures data. Statistics in Medicine, in preparation.
For further technical details please contact:
David WooffDepartment of Mathematical Sciences
University of Durham
Tel: +44 (0) 191 334 3121
Fax: +44 (0) 191 334 3051
Web: http://fourier.dur.ac.uk/
