17 Artificial intelligence for the safe use of medications during breastfeeding

<h3>Introduction</h3> Artificial intelligence is becoming a useful tool in clinical practice. We have previously published an explainable machine-learning algorithm for drug use in pregnancy based on multimodal data and suggest an orthogonal ensemble for modeling multimodal data. The model was trained with a set of labeled drugs and processed over 100,000 textual responses collected by our drug consultation service. Structured textual information is incorporated into the model by applying clustering analysis to textual features. Many medications that are not allowed during pregnancy are also not compatible with breastfeeding. Nevertheless, there are differences and there are medications that are not compatible with breastfeeding that are allowed during pregnancy, and vice versa. <h3>Objective</h3> The aim of the current study was to train the algorithm on lactation-related questions and to develop an algorithm for medication use during breastfeeding. <h3>Methods</h3> A multimodal machine learning models were developed for prediction of medication safety during lactation. We focus our efforts on tabular and molecule-related features to train ensemble-based machine learning models. The models were evaluated using a variety of evaluation schemes. <h3>Results</h3> The area under the receiver characteristic curve (AUC) of 0.921 for cross-validation applied to a dataset of 270 manually labeled drugs. An AUC of 0.963 was calculated when the algorithm was evaluated on a second, independent dataset, labeled by a second group of experts. On these two datasets, the highest performing model uses pregnancy safety and tabular, handcrafted drug features to predict lactation safety. Later, another dataset, consisting of 14 drugs, whose lactation safety is not aligned with pregnancy safety; an AUC of 0.906 was obtained on this dataset for a model trained with tabular, handcrafted, and molecule-based features only. For the later model, the most contributing features identified through SHAP analysis are antineoplastic agents (increase risk), anti-infective agents (decrease risk), and narrow therapeutic drugs (increase risk). Predictions for liraglutide (AUC=0.048) and rosuvastatin (AUC=0.401) for example, demonstrate the strengths and limitations of the model. <h3>Conclusions</h3> The models we developed can be used to make more informed decisions when a patient and her doctor discuss potential treatment options, thus improving the safety of drugs, saving lives, and reducing costs.