Exploring Transformer Model in Longitudinal Pharmacokinetic/Pharmacodynamic Analyses and Comparing with Alternative...

Area of Science:

• Pharmacometrics
• Quantitative Pharmacology
• Computational Biology

Background:

• Advances in machine learning have spurred interest in applying these techniques to quantitative pharmacology.
• Longitudinal pharmacokinetic/pharmacodynamic (PK/PD) modeling requires robust analytical methods to interpret complex time-series data.
• A comprehensive assessment of various natural language processing (NLP) algorithms for PK/PD analysis is needed.

Purpose of the Study:

• To investigate the application of the transformer model in longitudinal PK/PD data analysis.
• To compare the performance of different NLP models, including LSTM and neural-ODE, for PK/PD modeling.
• To evaluate model performance in predicting seen and unseen PK/PD regimens.

Main Methods:

• Utilized virtual PK/PD data across three distinct dosing regimens.
• Compared the predictive performance of transformer, long short-term memory (LSTM), and neural-ODE models.
• Assessed model accuracy for both training-included (seen) and training-excluded (unseen) regimens.

Main Results:

• LSTM and neural-ODE showed strong performance for seen regimens, with minor information loss for unseen regimens.
• The transformer model, similar to neural-ODE, excelled at describing time-series PK/PD data but struggled with precise extrapolation to unseen regimens.
• Incorporating a small amount of unseen data into the training set substantially improved predictive performance for all tested regimens.

Conclusions:

• This study pioneers the use of the transformer model for time-series PK/PD analysis.
• A systematic comparison reveals the strengths and limitations of current NLP models in PK/PD.
• Data augmentation with even minimal unseen data is a critical strategy to enhance predictive accuracy in PK/PD modeling.