Purpose: Develop a classificatory tool to identify different populations of patients with Traumatic Brain Injury based on the characteristics of deficit and response to treatment.
Work method: A KDD framework where first, descriptive statistics of every variable was done, data cleaning and selection of relevant variables. Then data was mined using a generalization of Clustering based on rules (CIBR), an hybrid AI and Statistics technique which combines inductive learning (AI) and clustering (Statistics). A prior Knowledge Base (KB) is considered to properly bias the clustering; semantic constraints implied by the KB hold in final clusters, guaranteeing interpretability of the resultis. A generalization (Exogenous Clustering based on rules, ECIBR) is presented, allowing to define the KB in terms of variables which will not be considered in the clustering process itself, to get more flexibility. Several tools as Class panel graph are introduced in the methodology to assist final interpretation.
Work results: A set of 5 classes was recommended by the system and interpretation permitted profiles labeling. From the medical point of view, composition of classes is well corresponding with different patterns of increasing level of response to rehabilitation treatments.
Discussion: All the patients initially assessable conform a single group. Severe impaired patients are subdivided in four profiles which clearly distinct response patterns. Particularly interesting the partial response profile, where patients could not improve executive functions.
Conclusions: Meaningful classes were obtained and, from a semantics point of view, the results were sensibly improved regarding classical clustering, according to our opinion that hybrid AI & Stats techniques are more powerful for KDD than pure ones.