EP-026 - MACHINE LEARNING FOR PHARMACOKINETIC/PHARMACODYNAMIC MODELING.

A. Tang; Thomas Jefferson High School for Science and Technology, Mclean, VA, USA.

Background: Recent efforts in applying machine learning (ML) for PK/PD modeling showed significant limitations on modeling/predicting sparse PK/PD data and extrapolations to unseen dose regimen and dose levels. In this work, a variety of new recurrent neural networks including the ODE-LSTM, Phased LSTM, CTGRU and GRU-D were evaluated using sparse and irregularly sampled PK/PD data and predicting PD data of unseen dosing regimens and dosing levels.

Methods: Sparse and irregularly sampled PK/PD data (6 and 12 time points per day) were simulated from the one-compartment absorption PK model and the Indirect PK/PD model with random errors in CL and PD measurements. Data of once daily (QD) with returning phases to baseline were used for developing ML models, whose performances of predicting PK/PD for twice daily (BID) and unseen dose levels were evaluated.

Results: All four methods were able to adequately extrapolate from QD to BID. The GRUD and Phased LSTM performed slightly better than the CTGRU and ODE-LSTM. There seemed no meaningful difference between the performance using 6 and 12 observations per day. Among the four methods, the GRUD was most accurate in predicting unseen dose levels, however, limited to only predict the PD for dose levels about 3-10 fold out of dose levels from training. The other methods were unable to predict unseen dose level. The ML models tested were able to recognize the hysteresis between PK and PD. The potential of over-fitting and under-fitting were also explored, which showed that the ODE-LSTM appeared to be most prone to these risks.

Conclusion: The new ML models seemed to overcome some limitations of previous works and showed promise of integrating neural networks in PK/PD. The GRUD method appeared to perform the best.