Predicting the need for CT imaging in children with minor head injury using an ensemble of Naive Bayes classifiers

OBJECTIVE Using an automatic data-driven approach, this paper develops a prediction model that achieves more balanced performance (in terms of sensitivity and specificity) than the Canadian Assessment of Tomography for Childhood Head Injury (CATCH) rule, when predicting the need for computed tomography (CT) imaging of children after a minor head injury. METHODS AND MATERIALS CT is widely considered an effective tool for evaluating patients with minor head trauma who have potentially suffered serious intracranial injury. However, its use poses possible harmful effects, particularly for children, due to exposure to radiation. Safety concerns, along with issues of cost and practice variability, have led to calls for the development of effective methods to decide when CT imaging is needed. Clinical decision rules represent such methods and are normally derived from the analysis of large prospectively collected patient data sets. The CATCH rule was created by a group of Canadian pediatric emergency physicians to support the decision of referring children with minor head injury to CT imaging. The goal of the CATCH rule was to maximize the sensitivity of predictions of potential intracranial lesion while keeping specificity at a reasonable level. After extensive analysis of the CATCH data set, characterized by severe class imbalance, and after a thorough evaluation of several data mining methods, we derived an ensemble of multiple Naive Bayes classifiers as the prediction model for CT imaging decisions. RESULTS In the first phase of the experiment we compared the proposed ensemble model to other ensemble models employing rule-, tree- and instance-based member classifiers. Our prediction model demonstrated the best performance in terms of AUC, G-mean and sensitivity measures. In the second phase, using a bootstrapping experiment similar to that reported by the CATCH investigators, we showed that the proposed ensemble model achieved a more balanced predictive performance than the CATCH rule with an average sensitivity of 82.8% and an average specificity of 74.4% (vs. 98.1% and 50.0% for the CATCH rule respectively). CONCLUSION Automatically derived prediction models cannot replace a physician's acumen. However, they help establish reference performance indicators for the purpose of developing clinical decision rules so the trade-off between prediction sensitivity and specificity is better understood.