Abstract
The management of proximal humeral fractures is challenging, and fixation plates often show a high failure rate. However, new fixation plates with variable angle screws could be beneficial. Finite element (FE) studies have shown some benefits of plates with variable angle screws, but not all possible combinations have been explored, and hence worst and optimal scenarios have not been identified. The full exploration of the solution space is not possible using FE techniques due to the computational expense; therefore, a more computationally affordable technique is needed. This study aimed to develop adaptive neural network (ANN) models that can predict the likelihood of a screw collision and the level of strain on the humeral bone when the orientation of the screws is changed. ANN models were trained using input and output data from FE simulations with varying screw angles, developed on a single subject with a two-part fracture in the proximal humerus. Training sets of different sizes were developed to determine the quantity of data required for an accurate model. Firstly, the ANNs were used to make predictions of results from FE unseen data, showing an 84.4% accuracy for the prediction of screw collision and good correlation (R(2) = 0.99) and low levels of error (RMSE = 0.65%-5.49% strain) for the prediction of bone strain. The ANNs were used to make predictions of a full factorial scenario, showing that the variation of the orientation of the screw in the calcar region has the greatest impact on the bone strain around all screws.