An artificial neural network-pharmacokinetic model and its interpretation using Shapley additive explanations

Area of Science:

• Pharmacokinetics and Pharmacodynamics
• Artificial Intelligence in Medicine
• Computational Biology

Background:

• Artificial neural networks (ANNs) face challenges in medical applications due to their "black-box" nature, hindering scientific confidence.
• Handling time-series data in pharmacometrics is a significant hurdle for ANN implementation.
• Existing population pharmacokinetic (PopPK) models may have limitations in accurately predicting complex pharmacokinetic profiles.

Purpose of the Study:

• To develop an interpretable ANN model for predicting time-series pharmacokinetics (PKs).
• To evaluate the scientific validity and predictive performance of the ANN-PK model against conventional PopPK models.
• To identify key factors influencing pharmacokinetic predictions using explainable AI techniques.

Main Methods:

• Developed an artificial neural network-pharmacokinetic (ANN-PK) model integrating patient data for clearance (CL) prediction.
• Utilized a one-compartment model with one-order absorption, updating ANN parameters via back-propagation.
• Applied Kernel SHAP to interpret the ANN-PK model and quantify input feature importance.

Main Results:

• The ANN-PK model demonstrated superior prediction accuracy with a root mean squared error (RMSE) of 31.0 ng/ml compared to the PopPK model's 41.1 ng/ml.
• Goodness-of-fit plots showed better convergence for the ANN-PK model, indicating improved performance.
• SHAP analysis identified age and body weight as the most influential covariates for CL prediction, which were then incorporated into the PopPK model.

Conclusions:

• The developed ANN-PK model effectively handles time-series pharmacokinetic data, offering higher prediction accuracy than conventional PopPK models.
• The integration of SHAP provides scientific validity to the ANN-PK model, enabling interpretation of predictions.
• This approach paves the way for developing interpretable AI models for predicting PKs, drug efficacy, and side effects in drug discovery and development.