Comparison of Classifier Algorithms in the Identification of Polypharmacy and Factors Affecting it in the Elderly Patients

Prescribing and consuming drugs more than necessary which is known as polypharmacy, is both waste of resources and harm to patients. Polypharmacy is especially important for elderly patients; therefore, the factors affecting it must be identified and analyzed properly.

In this retrospective study, first, several classifier algorithms, i.e., C4.5, SVM, KNN, MLP, and BN for polypharmacy identification were compared in terms of performance using WEKA software. In this process, 16 new features were extracted alongside the four existing features from data on 81,677 prescriptions of 19,428 outpatients aged 70 to 95 years whose prescriptions were dispensed in pharmacies contracted by the Iran Health Insurance Organization- Tehran province. The performance comparison was done using corrected t-test with resampling. In order to identify the effect of elderly patients’ characteristics on polypharmacy, two important parameters of the C4.5 were optimized by grid search using 50% of the dataset and then run on the rest of the dataset. The resulted rules were then presented in the form of a decision tree and verbal expressions.

Paired comparison of the classifiers indicated better performance of C4.5 and BN compared to the others. C4.5 had the ability to identify the factors that affect polypharmacy. In addition, parameter tuning improved the accuracy and AUC of applied algorithms. It also reduced the size of the resulted decision trees as well as the number of generated rules significantly.

The data mining approach and C4.5 can identify and explain the characteristics of the elderly effective on the polypharmacy. The higher percentage of visits to general practitioners and contacts with a limited number of pharmacies are the most important characteristics.