Drug-Drug Interactions (DDIs) Detection from On-Line Health Forums: Bi-Submodular Optimization (BSMO)

With the growth of mobile Internet, online health forums become more accessible for patient to health related discussions, subsequently host rich resources of drug-drug interactions (DDIs). However, traditional methods are not feasible for the large volume online data. They are designed for highly structured data sources such as clinical trials and spontaneous reporting systems, whose inherent limitations include low coverage and under-reporting. In this paper, we propose a bi-submodular optimization (BSMO) method to detect DDIs using the forum data collected online. The relationships between co-mentioned drugs and symptoms can be modeled with a conditional (predefined thresholds) graph, where a vertex represents either a drug or a symptom, and an edge represents the co-occurrence among drugs and/or symptoms. A connectedsub-graph consists of both symptom and drug vertexes reveals the occurrence of DDIs. A novel score function is proposed to characterize the degree of DDIs within a connected subgraph. Therefore the DDIs detection using on-line health forum data is then formulated as a sub-graph detection problem. An approximated algorithm was proposed based on bi-submodular optimization, then showed the complexity of the algorithm is nearly linear. Extensive experiments on the health forum data demonstrate the effectiveness and efficiency of our proposed approach.