Using multiple drug similarity networks to promote adverse drug event detection

Abstract

The occurrence of an adverse drug event (ADE) has become a serious social concern of public health. Early detection of ADEs can lower the risk of drug safety as well as the expense of the drug. While post-market spontaneous reports of ADEs remain a cornerstone of pharmacovigilance, most existing signal detection algorithms rely on substantial accumulated data, limiting their applicability to early ADE detection when reports are scarce. To address this issue, we propose a label propagation model for generating enhanced drug safety signals using multiple drug features. We first construct multiple drug similarity networks using a range of drug features. We then calculate initial drug safety signals using conventional signal detection algorithms. These original signals are subsequently propagated across each drug similarity network to obtain enhanced drug safety signals. We evaluate our proposed model using two common signal detection algorithms on data from the FDA Adverse Event Reporting System (FAERS). Results demonstrate that enhanced drug safety signals with pre-clinical information outperform the standard safety signal detection algorithms on early ADE detection. In addition, we systematically evaluate the performance of different drug similarities against different types of ADEs. Furthermore, we have developed a web interface (http://drug-drug-sim.aimedlab.net/) to display our multiple drug similarity scores, facilitating access to this valuable resource for drug safety monitoring.

Figure 1

Overall framework of label propagation using multiple similarity networks. The method includes three parts: (a) Original drug safety signal score computation, (b) Multiple similarity network construction, and (c) Enhanced drug safety signal score generation with similarity networks.

Figure 2

Frequency of no. of drugs associated with ADEs.

Figure 3

Frequency of no. of ADEs associated with drugs.

Figure 4

Comparison of the different methods based on MGPS on yearly cumulative reports.

Figure 5

Comparison of the different methods based on BCPNN on yearly cumulative reports.

Figure 6

Comparison of AUC achieved using the different methods based on MGPS on top 10 most frequent MedDRA SOCs. (#) at the end of SOC names in the y-axes denotes the number of unique MedDRA PTs from our filtered SIDER dataset mapped to each SOC.

Figure 7

Comparison of AUC achieved using the different methods based on BCPNN on top 10 most frequent MedDRA SOCs. (#) at the end of SOC names in the y-axes denotes the number of unique MedDRA PTs from our filtered SIDER dataset mapped to each SOC.

Figure 8

Landing page of the website.

Figure 9

Result page of the website.

Figure 10

Quarterly change in the rank of Pitavastatin-Rhabdomyolysis Drug-ADE pair.

Figure 11

Quarterly change in the rank of Nitroprusside-Hypotension Drug-ADE pair.