Injury Severity Analysis based on mutual information for in depth investigation of accident database

In Europe traffic accidents are now widely recorded in national databases. In view of the massive amounts of accident data, the use of data mining tools is essential to sift truly relevant information, and to extract reliable relationships between injury severity and potential causation factors. We present an innovative data mining approach for in depth investigation of causation in accidents databases. Classical statistical tools evaluate the strength of potential causal relationships by essentially linear techniques, or strongly rely on ad hoc specific models. We outline here how mutual information ratios (based on conditional entropies) contribute to rigorously quantify the influence of causation factors on accident outcome descriptors such as injury type and severity. Information theoretic methods help to automatically select small groups of factors with high causation impact on accidents severity, with no hypothesis on underlying relationships between observed variables. We successfully apply this approach to analyze causation factors in the German In Depth Accident Study database, which is one of the largest and most complete in depth accident survey and data collection in Europe.