The Temporal Discovery Workbench (TDWB) was first developed to discover associations / rules in temporal datasets where occasional unexpected events occur. This has been applied so far to the discovery of myocardial damage in ICU patient datasets. Subsequently, we have added a module which describes temporal data in terms of descriptor-ranges and the level of correlation between 2 designated descriptors. The Correlation module has been applied to datasets from traumatic brain injury patients, and the correlation reported in this study is between PbtO2 (measurement of oxygenation of cerebral cells) and CPP (Central Perfusion Pressure); also the values of descriptors CPP, PaO2 and PbtO2 are reported. This analysis has shown that many of the variables remain quite stable for reasonable periods of time and we now refer to such periods as (stable) sequences. In these Knowledge Acquisition sessions, the expert identified important descriptors (CPP, PaO2, PbtO2 & ORx/CORR), their ranges and the boundaries for these ranges. Subsequently, he also identified Clinical Actions which should be applied to the sequences encountered, and to additional sequences which can occur. After some manual analysis, the expert provided us with a preliminary set of rules which relate Descriptor-Value pairs and Clinical Actions. An annotation capability has since been added to TDWB's correlation module. The domain expert, JR, also added these rules to a spreadsheet and we have shown that the 2 analyses produce virtually identical Clinical Actions.
Even though we already have an initial (approximate) set of rules provided by the domain expert, we decided to determine how these might be revised / refined when the annotated instances were processed by an appropriate Machine Learning algorithm. (This was also a prelude to applying this overall approach to other areas of (critical care) medicine.) We have since run a series of studies using a Decision Tree algorithm which has produced strong agreements between the annotations produced by the expert's informal ruleset and the categories suggested by the Decision Tree algorithm. These results will be discussed in some detail, as will our planned future work which includes using this approach with "live" patients to direct clinical care.
(Joint work with Samuel Cauvin (Department of Computing Science, University of Aberdeen) and Jonathan Rhodes (Intensive Care Unit, Western General Hospital, Edinburgh)
Host: Jessica Chen-Burger