Machine learning approaches for predicting the onset time of the adverse drug events in oncology

Predicting the onset time of adverse drug events can substantially lessen the negative impact on the prognosis of cancer patients who are often subject of aggressive and highly toxic treatment regimens. However, the laboratory verification of each patient case to study the mechanics of adverse drug events requires costly, time-intensive research. Thus, to alleviate the efforts required to tackle this problem, using computational models is highly desirable. To provide a suite of such applicable models, we used openly available adverse drug event data resources called FAERS and explored various machine learning paradigms to assess their performance in predicting adverse effect onset days (since the beginning of the treatment). Among various machine learning approaches, we observed that the graph-based embedding model, particularly ComplEx, performed better than other, more traditional machine learning approaches. The embedding learned from the ComplEX trained with k-NN regression for the downstream predictive task obtained the lowest root mean square error, which we consider very promising for further research.