Predicting Adverse Effects of Drugs

​

     The main objective of this dissertation was to assess how useful Data Mining methods and algorithms could predict the adverse effect of drugs. Thus, two experiences were performed. Data set was taken from the ADReCS database.

      In the first experiment, were used recommendation algorithms: MF, Slope One and User k-NN. The algorithm that obtained a model with better performance was the MF, since this algorithm obtained the value of the greater Accuracy metric (about 45.15%). However, the results obtained were not promising.

    In the second experiment, we have defined a classification tasks using only the top of the ADRs identifiers hierarchy. The classification algorithms used were Decision tree, Random Forest, Naive Bayes and SVM. Considering the results obtained, it can be stated that the algorithm that obtained a model with better performance was the SVM, since this algorithm obtained the best value of Accuracy (about 94.41%). The results obtained were satisfactory.
     After analyzing the results achieved in the two experiments, it was observed that experiment 2 obtained better results. Thus, it is concluded that making individual predictions of adverse effects is quite complex, and when reducing the detail of the information, that is, when one goes up the level of hierarchy, the obtained results are better. It is important to note that this improvement of results also occurs when done with the Feature Selection, since a pre-selection of the best attributes occurs. In general, this results were promising and encourage the continuation of this line of research.

​