Interpretable analysis of NARX neural network prediction MPA-AUC

Clinically, the concentration of Mycophenolic Acid (MPA) within 12 hours after kidney transplantation is usually monitored for immune rejection, and the Area Under the Curve of MPA (MPA-AUC0-12h) is measured. Current methods usually select plasma concentration by Limited Sampling Strategy (LSS), and use Multiple Linear Regression (MLR) or Artificial Neural Network to predict MPA-AUC0-12h. MLR only predicts one administration regimen, while other administration regimens need to recalculate the model, which will greatly increase the burden on clinicians, and the generalization ability of the model is weak. Other models often ignore the time series between plasma concentrations, and the accuracy of prediction is low and the error is large, so they are not capable of clinical generalization. In addition, these models do not carry out interpretability analysis on the prediction results, and there is a lack of interpretability for the prediction of the black box model. We propose an Interpretable Analysis of the Nonlinear Autoregressive with External Input Neural Network (IA-NARX) model. Our model combines with the time series of patient's plasma concentration, uses NARX model to predict MPA-AUC0-12h, and uses Shapley Additive ex-Planations (SHAP) to interpretability analysis on the prediction results. In conclusion, the experimental results show that the accuracy of our model is 82.14%, Mean Absolute Error (MAE) is 9.863, Mean Absolute Percentage Error (MAPE) is 0.1872, Root Mean Square Error (RMSE) is 12.232. The prediction effect of IA-NARX model is superior to existing models in terms of accuracy and precision. In addition, the interpretability of our model is in line with clinical expectations, and it has a good prospect of popularization and application.