Using dependency/association rules to find indications for computed tomography in a head trauma dataset.

Analysis of a clinical head trauma dataset was aided by the use of a new, binary-based data mining technique, termed Boolean analyzer (BA), which finds dependency/association rules. With initial guidance from a domain user or domain expert, the BA algorithm is given one or more metrics to partition the entire dataset. The weighted rules are in the form of Boolean expressions. To augment the analysis of the rules produced, we applied a probabilistic interestingness measure (PIM) to order the generated rules based on event dependency, where events are combinations of primed and unprimed variables. Interpretation of the dependency rules generated on the clinical head trauma data resulted in a set of criteria that identified minor head trauma patients needing computed tomography (CT) scans. The BA criteria contained fewer variables than were found using recursive partitioning of Chi-square values (five variables versus seven variables, respectively). The BA five-variable criteria set was more sensitive but less specific than the seven-variable Chi-square criteria set. We believe that the BA method has broad applicability in the medical domain, and hope that this paper will stimulate other creative applications of the technique.