Explainable multimodal machine learning model for classifying pregnancy drug safety.

Area of Science:

• Computational biology
• Pharmacology
• Machine learning

Background:

• Teratogenic drugs pose significant risks to fetal development, but safety data is lacking for many approved medications.
• Accurate prediction of drug safety during pregnancy is crucial for maternal and fetal health.
• Existing methods for assessing teratogenic risk are limited, necessitating advanced predictive approaches.

Purpose of the Study:

• To develop and validate an explainable machine learning model for classifying drug safety during pregnancy.
• To assess the teratogenic risks of specific drugs, including Varenicline, Mebeverine, and Meloxicam.
• To provide a tool for physicians to evaluate drug risks during pregnancy.

Main Methods:

• Created a labeled drug dataset by processing over 100,000 textual responses from a teratology information service.
• Applied clustering analysis to textual features for incorporating structured information into the model.
• Developed an orthogonal ensemble model for multimodal data integration and classification.
• Utilized cross-validation and cross-expert validation for model performance assessment.

Main Results:

• Achieved an area under the receiver operating characteristic curve (AUC) of 0.891 with cross-validation and 0.904 with cross-expert validation.
• Identified Varenicline and Mebeverine as likely safe for pregnancy.
• Indicated a higher risk associated with Meloxicam, a nonsteroidal anti-inflammatory drug (NSAID), during pregnancy.
• Developed a web-based application for physicians to assess drug risks.

Conclusions:

• The proposed explainable machine learning model effectively predicts pregnancy drug safety.
• The model provides valuable insights into the risks of specific drugs, aiding clinical decision-making.
• The developed tool can enhance the management of medication use during pregnancy.