2021 FDA Science Forum
Long-Short Term Memory Recurrent Neural Network for Pharmacokinetic-Pharmacodynamic Modeling
- Authors:
- Center:
-
Contributing OfficeCenter for Drug Evaluation and Research
Abstract
Background
The traditional compartment modeling methods require specification of model structures based on simplified hypotheses that may not represent complex pharmacokinetics (PK) and pharmacodynamics (PD) accurately. Recurrent neural network (RNN), one of the popular machine learning methods, has been demonstrated as a powerful tool that can learn from various types of time series data and make appropriate predictions without prior knowledge.
Purpose
This project was designed to explore the ability of the Long Short-Term Memory (LSTM) RNN network to analyze PK/PD data and predict new scenarios.
Methodology
A hypothetical orally administrated drug X was used to generate PK and PD profiles for 60 simulated patients. The PK and PD of drug X is described by a three-compartment population PK model and an indirect-response model. Each patient was assumed to take doses (250-2000 mg) under various dosing regimens (i.e. QD, BID, TID, and random) for 7 days. The measurements of plasma concentration and PD effects were taken every 0.5 hours. The training process for LSTM was performed using the time sequence of plasma concentration of the drug, dose, and subject demographic information under QD regimen as inputs/features and the PD response variable as the output/target. The time series under BID, TID, and random-dosing regimens were considered as testing datasets for evaluating the extrapolating ability of the proposed model. Furthermore, a sparse subset with 100 randomly selected measurement points for each dose was generated to investigate the performance of the LSTM network for modeling data with irregular time intervals.
Results
The optimized LSTM model captured temporal dependencies and predicted PD profiles accurately for the simulated indirect PK-PD relationship. The model trained from relatively sparse data shows an inferior predicting performance, especially when predicting the PD outcomes under other regimens.
Conclusion
Generic LSTM models can approximate the complex physiologic PK/PD relationship. Machine learning models and mechanistic models can certainly complement each other to help improve the efficiency of drug development and optimize the treatment for individual patients.
