Novel Hybrid Artificial Neural Network-Nonlinear Mixed Effect Model Modeling Approach for Population Data Analysis in Model-based Drug Development
Chee M Ng
Children Hospital of Philadelphia and School of Medicine, University of Pennsylvania, Philadelphia, PA
Background and Objectives: A potential limitation of current model-based drug development (MBDD) approach is that integration of the knowledge from observed data into population model can be incomplete (incomplete learning) due to computing hardware/software limitation and lack of biological understandings. This resulted in biased model prediction that may lead to inefficient drug development strategy. Artificial neural network (ANN) is a powerful function mapping tool and any well-behaved multivariate functional relationships can be implemented exactly with ANN. Therefore, it is hypothesized that ANN can be used to detect/correct the deficiency of the population model developed with nonlinear mixed-effect model (NLME) and improve the model prediction throughout the MBDD process. The objective of this study is to develop a novel hybrid ANN-NLME modeling approach for population data analysis in MBDD.
Methods: The developed ANN-NLME method consists of a three-layer fully connected feed-forward ANN with Bayesian regularization and NONMEM FOCE. A MATLAB program was written to integrate the ANN with NONMEM for data analysis. Three simulated dataset/scenarios were used to assess the performance of the ANN-NLME (FOCE) (Table 1). The complete model was used to simulate population data. Three different models including 1) complete model with NONMEM FOCE (C), 2) Incomplete model with NONMEM FOCE (IC), and 3) Incomplete model with ANN-FOCE (IC-ANN), was fitted to the simulated data and then results were compared.
Table 1: Scenarios for Model Comparison
|
Scenarios |
Complete Model |
Incomplete Model |
|
A |
2-compartment PK |
1-compartment PK |
|
B |
2-compartment PK IgG model with time-dependent inhibitory effect of anti-IgG antibody on PK of the low dose group |
2-compartment PK |
|
C |
PK-PD model with two interacting PD pathways (I and II) |
PK-PD model with a pathway I |
Results/Conclusion: The ANN was able to detect/correct the model deficiency and improve the prediction of the model with incomplete information (Table2). To my best knowledge, this is the first reported hybrid ANN-NLME modeling approach for population data analysis. This novel approach that combines the powerful mapping function of ANN with flexibility of NLME method may serve as an excellent computational platform for developing highly predictive population model to support decision making in MBDD.
Table 2. RMSE of Model Prediction. (Lower=Better)
|
Scenarios |
IC |
IC-ANN |
C |
|
A |
0.223 |
0.104 |
0.094 |
|
B |
0.373 |
0.178 |
0.095 |
|
C |
0.238 |
0.119 |
0.095 |
References:
[1] Kolmogorov A. On the representation of continuous functions of several variables by superposition of continuous functions of one variable and addition. Doklady Akademiia Nauk SSSR 1957;114:953-6
[2] McKay DJC. Bayesian interplotation. Neural Computation 1992;4:415-47