Efficacy of predictions of post-operative gait in rectus transfer patients using neural networks

Introduction. Cerebral palsy patients exhibit a wide range of functional responses to specific surgical interventions. The variability associated with surgical outcomes within this patient population precipitates the need to determine the functional effects in advance of the surgery. This study seeks to determine the accuracy and stability of neural network models for the prediction of post-surgical kinematics in order to provide the surgeon with important information during the pre-surgical decision-making process.; Methods. Post-surgical kinematic outcomes were predicted using a variety of neural network and statistical techniques. Several methods were used in the predictions. Preliminary studies were completed using the Ward Backprop networks with twenty-three legs. However, this network did not prove to be as accurate as the Cascade Correlation network, and so it was not used in the final analysis. Another study used 100 legs to compare Cascade Correlation networks with locally weighted regression, a statistical matching and prediction technique. Finally a new method that used distance metrics to pre-sort data for inclusion in the neural network training set was evaluated.; The similarity of the predictions to the post-surgical data was evaluated using statistical similarity measures including a distance metric, and a clinical similarity statistic. This latter statistic was derived from a web based survey of gait clinicians.; Results and discussion. The neural network model was found to extrapolate predictions more accurately than shape analysis with locally weighted regression. Results indicated that it was possible to predict 60% of the post-surgical waveforms for the hip and 75% of post-surgical waveforms for the knee using the neural network compared to 35% and 60% respectively using locally weighted regression. Thirty percent of neural network predictions were within the average daily variation of knee flexion found by Miller et al. (1996). Where prediction was not possible, pre-surgical characteristics that may confound the predictive models were identified. These included subjects who did not exhibit impact absorption during support phase, and subjects with crouched gait. Finally, the new pre-sorting technique appeared to be on a par or better than earlier techniques in predicting post-surgical outcomes.