A dynamic prediction engine to prevent chemotherapy-induced nausea and vomiting

Background Cancer remains the second major cause of death in the United States over the last decade. Chemotherapy is a core component of nearly every cancer treatment plan. Chemotherapy-Induced Nausea and Vomiting (CINV) are the two most dreadful and unpleasant side-effects of chemotherapy for cancer patients. Several patient-specific factors affect the risk of CINV. However, none of the guidelines consider those factors. Not all of the patients have the similar emetic risk of CINV. Despite the improvements in CINV management, as many as two-thirds of chemotherapy patients still experience some degree of CINV. As a result, physicians use their personal experiences for CINV treatment, which leads to inconsistent managements of CINV. Objective The overall objective of this study is to improve the prevention of CINV using precise, personalized and evidence-based antiemetic treatment before chemotherapy. In CINV prediction, one of the interesting factors is that CINV has two distinct and complex pathophysiologic phases: acute and delayed. In addition, the risk factors and their associations are different for different emetogenic chemotherapies (e.g., low, moderate, and high). There are six contexts considering the combination of phases and emetogenicity levels. This will require the creation of six different models. Instead, our objective was to describe a single framework named “prediction engine” that can perform prediction query without losing the sensitivity to each context. The prediction engine discovers how the patient-related variables and the emetogenecity of chemotherapy are associated with the risk of CINV for each phase. Methods This was a single-center retrospective study. The data were collected by retrospective record review from the electronic medical record system used at the University of Missouri Ellis Fischel Cancer Center. An association rule-based dynamic and context-sensitive Prediction Engine has been developed. Physicians receive feedback about CINV risks of patients from the CINV decision support system based on patient-specific factors. Results The prediction performance of the system outperformed many popular prediction methods and all the results of CINV risk prediction published in the literature. Best prediction performance was achieved using the rule-ranking approach. The accuracy, sensitivity, and specificity were 87.85 %, 87.54 %, and 88.2 %, respectively. Conclusions The system used the patient-specific risk factors for making personalized treatment recommendations for CINV. It solved a real clinical problem that will shorten the gap between clinical practices and evidence-based guidelines for CINV management leading to the practice of personalized and precise treatment recommendation, better life quality of patient, and reduced healthcare cost. The approach presented in this article can be applied to any other clinical predictions.