110: Predicting Pregnancy Population Pharmacokinetics Using Machine Learning Algorithms

Statement of Purpose: Pregnant women are often an understudied population due to historically being excluded from clinical trials. Because many physiological changes occur during pregnancy some drug pharmacokinetics (PK) may differ and make it necessary for dose adjustments to ensure safety and efficacy. Artificial intelligence (AI) for predicting drug PK is a relatively new and emerging tool. Several papers have been published covering various methods for using AI machine learning (ML) to predict PK. The objective of this study is to determine which ML method will best predict the plasma concentration of simulated pregnant subjects taking various doses of labetalol. In the following experiment, we used various machine learning algorithms to determine which algorithm best predicted PK data in pregnant women.

Description of Methods & Materials: Population plasma PK data for pregnant women was generated using a physiologically based pharmacokinetic model with 10-30% variation in parameters with Berkeley Madonna. The time since the first dose was given, time since last dose, dose frequency, dose amount, study number, patient number and dosing cycle was also incorporated into the datafile for each patient. The data for each patient was appended together to represent a single study. Fifty subjects were simulated for each study. Multiple scenarios in which dose (200-1,000 mg), and frequency (6-12 hours) were varied for each study. The data file was randomized for training, validation, and test sets for the AI models. Once the data set was split using python-based software, the performance was evaluated using the coefficient of determination (R²) for the regression models and the root mean squared error (RMSE) for predictive error.

Data & Results: The initial results showed that linear regression and random forest models were greater than 0.6 R². But the mean squared error (MSE) was exceptionally high. The multivariate polynomial regressor model had an R² of greater than 0.7 and a lower MSE than either of the previous two algorithms.

Interpretation, Conclusion or Significance: The ML models which performed better than the others tested had an R² value of greater than 0.6. The MSE and R² improved with a multivariate polynomial regressor model vs its counterparts. However, the RMSE overall needed improvement. The results will guide further exploration of other model features for the best performing ML methods and algorithms. The application of ML use in pregnancy can bring increased discussion about the scarcity of data in this area which requires hundreds to thousands of data points.

Disclosures: The information in these materials is not a formal dissemination of information by FDA and does not represent agency position or policy.