(T1330-02-08) Machine Learning Approaches for Assessing Medication Transfer to Human Breast Milk: Estimating Milk/Plasma Drug Ratios

Purpose: The measurement of the human milk/plasma (M/P) drug concentration ratio holds critical significance in clinical pharmacology, particularly for breastfeeding mothers undergoing medical treatment. This ratio is a crucial determinant of the extent to which drugs ingested by the mother can pass into breast milk and consequently reach the nursing infant. Ensuring the medication levels remain within safe limits is vital not only for the health of the infant, who might be exposed to the drugs through breastfeeding, but also to maintain the efficacy of treatment for the mother. This study utilizes machine learning (ML) techniques such as K-Nearest Neighbors (KNN), Random Forest, Support Vector Machine (SVM), and Neural Networks to create a robust predictive model for M/P ratios within certain intervals. The proposed ML predictive models will enhance the safe administration of medications for nursing mothers, even in cases where in-depth understanding of the complex biological mechanisms that govern drug transfer into breast milk is not readily available.

Methods: The study's methodological framework incorporates a stratified approach to categorize the M/P ratios into distinct intervals, optimizing the predictive accuracy of the ML models. The M/P ratios are first dichotomized into lower transfer potential for the intervals of (0 to < 1) and higher transfer potential for (≥1). Additionally, a refined three-category system further discriminates these intervals into (0 to < 0.5), (0.5 to < 1), and (≥1), allowing for more granularity and precision. The research employed an extensive dataset, assembling a total of 162 drugs along with 11 predictor variables from literature. This rich dataset serves as the foundation for training a suite of ML models, each evaluated through 5-fold cross-validation techniques to ensure robustness and minimize overfitting. Unsupervised techniques such as Principal Component Analysis (PCA) were also utilized to reduce the feature space to its most informative axes, while the Bayesian Information Criterion (BIC) was used in model selection, ensuring the balance between model complexity and explanatory power. Mean predictive accuracy was used to return the best performed model.

Results: Among the four ML methods, KNN models achieved an accuracy of 83.7% for the dichotomized M/P ratios and 71.4% for the three-category system. The Random Forest models exhibited accuracies of 79.6% and 67.3% respectively, while the SVM approach yielded similar performance with 81.6% and 67.3% accuracy rates. Neural Networks, leveraging their capability to model non-linear relationships, displayed accuracies of 83.7% and 69.4% for the two categorizations, respectively. All models working with dichotomized M/P ratios yielded acceptable accuracy rates.

Conclusion: This work demonstrates that Machine Learning is a promising approach for predicting M/P ratios, particularly for two-category predictions. These techniques can offer valuable insights in risk assessment for compounds under early-stage development before the acquisition of deep mechanistic insight. However, the accuracy for three-category predictions is < 70%, highlighting the necessity of training models with a larger dataset for complex classifications.

References: Yang, H., Xue, I., Gu, Q., Zou, P., Zhang, T., Lu, Y., Fisher, J., & Tran, D. (2022). Developing an In Vitro to In Vivo Extrapolation (IVIVE) Model to Predict Human Milk-to-Plasma Drug Concentration Ratios. Molecular Pharmaceutics, 19(7), 2506-2517.
Fop, M., & Murphy, T. B. (2018). Variable selection methods for model-based clustering. Statistics Surveys, 12, 18-65.

Raftery, A. E., & Dean, N. (2012). Variable Selection for Model-Based Clustering. Journal of the American Statistical Association, 101(473), 168-178.
Stratigakis, A., Paty, D., Zou, P., Zhao, Z., Li, Y., & Zhang, T. (2023). A regression approach for assessing large molecular drug concentration in breast milk. Reproduction and Breeding, 3(4), 199-207.
Keutzer, L., You, H., Farnoud, A., Nyberg, J., Wicha, S. G., Maher-Edwards, G., Vlasakakis, G., Moghaddam, G. K., Svensson, E. M., Menden, M. P., Simonsson, U. S. H., & On Behalf Of The Unite Tb Consortium. (2022). Machine Learning and Pharmacometrics for Prediction of Pharmacokinetic Data: Differences, Similarities, and Challenges Illustrated with Rifampicin. Pharmaceutics, 14(8), 1530.
Zhao, C., Zhang, H., Zhang, X., Zhang, R., Luan, F., Liu, M., Hu, Z., & Fan, B. (2006). Prediction of milk/plasma drug concentration (M/P) ratio using support vector machine (SVM) method. Pharmaceutical Research, 23(1), 41–48.

Acknowledgements: Funding support: This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Disclosures of conflicts of interest: All authors declare no conflicts of interest relevant to the content of this article.


Figure 1: Paired plot of data with 3 categories of M/P ratio


Figure 2: The prediction accuracy matrix of KNN with 3 categories of M/P ratio


Figure 3: The table of data