A decision tree based neural network method for prediction of poor prognosis in traumatic brain injury patients

This paper aims to predict poor prognosis in traumatic brain injury(TBI) patients based on admission findings. A neural network is mapped from an initial decision tree constructed from a part of data. Then the designed network is trained and validated by the remained data. 10-fold cross validation method is applied and the area under the ROC curve and accuracy rate are reported. The most important attributes are determined from the trained network by two methods; Change of Mean of Squared Error (COM) and Sensitivity Analysis (SA). The results reveal a high accuracy rate (91.1%) and significant area under the ROC curve (0.655) for the model used in this study. Furthermore, the order of the important attributes in prediction which has determined by this method is accepted in clinical aspect. Accordingly, the combination of different modeling method may bring out some complexities in computations and interpretations but improves the accuracy in prediction.

Keywords:

Poor prognosis , Traumatic brain injury patients , Neural network , Decision tree