[4F2] Load prediction using an intraoperative joint sensor and artificial neural network
S Al-Nasser¹, S Noroozi¹, R Haratian¹, N Aslani² and A Harvey²
¹Bournemouth University, UK
²The Royal Bournemouth Hospital, UK
Predicting the load from sensors for intraoperative joint assessment using artificial neural networks (ANNs) is a relatively new and novel technique that can be useful for various other health monitoring applications. In joint replacement surgeries, balancing the load using ANN predicted load measurements can prove useful in cases where a larger sensing area is required and the geometry of the sensor is more complicated. Once trained, ANNs can provide real-time load predictions accurately and precisely; however, finding the optimal combination of parameters and hyperparameters to yield the most accurate results requires exploration.
In this research, training data was collected by loading the sensor with known weights. Then, the dataset was systematically preprocessed and trained using an ANN with different combinations of parameters and hyperparameters to investigate its performance based on different configurations of an ANN. It was found that the Levenberg-Marquardt back-propagation function, with 50 hidden layers and a learning rate of 0.1, yielded the lowest mean square error (MSE) of 0.0034 ± 0.000141, which has also been the best performing training function in other studies. However, the time required to perform this function was significantly longer than for any other combinations (P < 0.001). Additionally, lowering the learning rate for the Levenberg-Marquardt back-propagation function and the Broyden-Fletcher-Goldfarb-Shanno (BFGS) quasi-Newton back-propagation function resulted in no change in the performance.
In summation, this study accentuates the importance of a methodical approach in finding the optimal combination of parameters and hyperparameters for ANNs to achieve an adequate performance. It also proves that ANNs can aid in predicting the load from sensors, especially in cases where non-linearity is observed. Finally, translating the performance of the network to real-time data collection presents new challenges that must be addressed in the future. However, once the ANN is trained and validated, load balancing during joint replacements can be performed with more accuracy and precision.
In this research, training data was collected by loading the sensor with known weights. Then, the dataset was systematically preprocessed and trained using an ANN with different combinations of parameters and hyperparameters to investigate its performance based on different configurations of an ANN. It was found that the Levenberg-Marquardt back-propagation function, with 50 hidden layers and a learning rate of 0.1, yielded the lowest mean square error (MSE) of 0.0034 ± 0.000141, which has also been the best performing training function in other studies. However, the time required to perform this function was significantly longer than for any other combinations (P < 0.001). Additionally, lowering the learning rate for the Levenberg-Marquardt back-propagation function and the Broyden-Fletcher-Goldfarb-Shanno (BFGS) quasi-Newton back-propagation function resulted in no change in the performance.
In summation, this study accentuates the importance of a methodical approach in finding the optimal combination of parameters and hyperparameters for ANNs to achieve an adequate performance. It also proves that ANNs can aid in predicting the load from sensors, especially in cases where non-linearity is observed. Finally, translating the performance of the network to real-time data collection presents new challenges that must be addressed in the future. However, once the ANN is trained and validated, load balancing during joint replacements can be performed with more accuracy and precision.
