Introduction of an artificial neural network-based method for concentration-time predictions

Area of Science:

• Pharmacology
• Computational Biology
• Machine Learning

Background:

• Population pharmacokinetic (PK) modeling is essential in clinical pharmacology but is resource-intensive and requires specialized expertise.
• Traditional PK methods have seen limited evolution despite their proven utility.
• There is a need for more efficient and accessible methods for predicting drug concentration-time profiles.

Purpose of the Study:

• To investigate artificial neural networks (ANNs) as a supplementary approach to traditional population PK modeling.
• To assess the ability of ANNs to predict concentration-time curves using simulated and real clinical data.
• To evaluate the application of transfer learning for adapting ANN predictions to new patient populations and dosing schemes.

Main Methods:

• Utilized simulated data to train ANNs, bypassing the need for extensive clinical datasets.
• Designed a pharmacologically informed network architecture to enhance extrapolation capabilities.
• Employed transfer learning to adapt trained ANNs to different patient groups (e.g., hepatic impairment) and dosing regimens.
• Validated ANN predictions against real clinical data.

Main Results:

• ANNs successfully learned the dynamics of concentration-time curves and made accurate individual predictions from limited PK data.
• An ANN trained on simulated data demonstrated effective extrapolation to different dosing schemes when applied to real clinical data.
• Transfer learning enabled efficient adaptation of a healthy subject-trained ANN to predict PK in simulated hepatic impaired patients.
• The study confirmed ANNs' capability to learn PK profiles and adapt to new scenarios.

Conclusions:

• Artificial neural networks present a viable and efficient alternative for individual concentration-time predictions within pharmacokinetic workflows.
• ANNs can supplement established methods, offering faster and potentially more accessible predictions.
• Further research into the limitations and advantages of ANN-based PK modeling is warranted to optimize clinical application.