Artificial neural network predicts CT scan abnormalities in pediatric patients with closed head injury.

BACKGROUND

Artificial neural networks (ANNs) use nonlinear statistical modeling techniques to explore relationships in complex clinical situations. This study compared predictive ability of a trained ANN model to that of physician prediction of cranial computed tomographic (CT) scan abnormalities in children with head injury.

METHODS

A prospective cohort of 351 patients who presented with head trauma and underwent CT scans were studied. All pertinent data on historical and demographic information, and clinical features were recorded. Emergency department physicians used clinical judgment to record pretest probability of abnormal CT scans for all patients prospectively. Similar data from a retrospective chart review of 382 patients with head injury in the immediate preceding year were collected and used to train the ANN. Data from the prospective study was used to validate the ANN, construct a logistic regression model, and compare physician prediction.

RESULTS

Forty-five (12.9%) of 351 patients had abnormal CT scans. In predicting CT scan abnormality, the ANN model was more sensitive (82.2%) compared with physician prediction (62.2%).

CONCLUSION

ANNs may serve as a useful aid for decision support for emergency physicians in predicting intracranial abnormalities in closed head injury.