AI-001 - APPLICATION OF GAUSSIAN PROCESS MODEL IN LONGITUDINAL PHARMACOKINETIC/PHARMACODYNAMIC DATA AND COMPARISON WITH NATURAL LANGUAGE PROCESSING MODELS

Y. Cheng, X. Hao, H. Hu, Y. Li; Bristol Myers Squibb.

Background: Analyzing longitudinal pharmacokinetic/pharmacodynamic (PK/PD) data using natural language processing (NLP) model has received great attention recently, including long short-term memory (LSTM) and neural-ODE. However, one inherent feature of these models is that they are constructed with high volume of parameters, which inevitably renders simple data (i.e., PK/PD) fit into complicated model structure. To address this, we applied a Bayesian non-parametric modelling approach, known as Gaussian process (GP), to characterize the longitudinal PK/PD relationship. Comparison of GP against published NLP models was also conducted.

Methods: Simulated data (N=600) based on published population PK/PD model were randomly split to train/validation/test datasets. Input features include dose amount, demographics, and PK at specified time points. Outputs are the PD at specified time points. Root mean square error (RMSE) and overlaid plot between PD predictions and observations were assessed. The model performance in extrapolating data from unseen regimens (not in the training dataset) was evaluated.

Results: The GP model was trained using data from TID regimen and showed good performance in describing TID data (RMSE: 0.833). When extrapolating to unseen regimens, GP accurately predicted data for BID and QD regimens, with RMSE 1.507 and 2.536, respectively. All three regimens showed good concordance between prediction and observation. However, while reasonable accuracy was noted for LSTM (RMSE: 1.623, 3.335 and 9.678 for TID, BID and QD respectively) and neural-ODE (RMSE: 1.076, 1.887 and 4.575 for TID, BID and QD respectively) in similar setting, extrapolating to unseen regimens led to loss of information in capturing the data pattern, including failure to mirror the magnitude of data fluctuation (i.e., trough to max) when examining the overlaid plot.

Conclusion: The analysis represents the first work to apply GP model and benchmark with NLP models in longitudinal PK/PD space. Within the scope of our work, GP outperformed NLP models in the context of better prediction accuracy, particularly when extrapolating to unseen regimens as well as allowing probabilistic prediction. The finding suggested that GP, as a non-parametric machine learning approach is a promising direction for future investigation.