Mining Patient-Generated Content for Medication Relations and Transition Network to Predict the Rankings and Volumes of Different Medications

Accurate estimates of medication rankings and volumes can benefit patients, physicians, online health communities, pharmaceutical companies, and the healthcare industry at large. This study analyzes patient-generated content in online health communities to discover important medication transition and combination patterns for better ranking and volume predictions. The current research takes a data-driven analytics approach to identify medication information from patient posts and classify different types of medication relations. The identified relation patterns then are represented in a medication relation network with an adjusted transition model for ranking and volume estimates. Evaluation results of real-world patient posts show the proposed method is more effective for predicting medication rankings than existing network-based measures. Moreover, a regression-based model, informed by the proposed method’s network-based outcomes, attains promising accuracy in estimating medication volumes, as revealed by the relatively low mean squared errors. Overall, the proposed method is capable of identifying important features for increased predictive power, as manifested by $${\text{R}}^{2}$$ R 2 and adjusted $${\text{R}}^{2}$$ R 2 values. It has the potential for better medication ranking and volume predictions, and offers insights for decision making by different stakeholders. This method is generalizable and can be applied in important prediction tasks in healthcare and other domains.