Development of a prediction models for chemotherapy-induced adverse drug reactions: A retrospective observational study using electronic health records

Purpose: Chemotherapy-induced adverse drug reactions (ADRs) are common and diverse, and not only affect changes or interruptions to treatment schedules, but also negatively affect the patient's quality of life. This study aimed to predict eight chemotherapy-induced ADRs based on electronic health records (EHR) data using machine-learning algorithms. Methods: We used EHR data of 6812 chemotherapy cycles for 935 adult patients receiving four different chemotherapy regimens (FOLFOX, 5-fluorouracil + oxaliplatin + leucovorin; FOLFIRI, 5-fluorouracil + irinotecan + leucovorin; paclitaxel; and GP, gemcitabine + cisplatin) at a tertiary teaching hospital between January 2015 and June 2016. The predicted ADRs included nausea-vomiting, fatigue-anorexia, diarrhea, peripheral neuropathy, hypersensitivity, stomatitis, hand-foot syndrome, and constipation. Three machine learning algorithms were used to developed prediction models: logistic regression, decision tree, and artificial neural network. We compared the performance of the models with area of under the ROC (Receiver Operating Characteristic) curve (AUC) and accuracy. Results: The AUCs of the logistic regression, decision tree, and artificial neural network models were 0.62-0.83, 0.61-0.83, and 0.62-0.83, respectively, and the accuracies were 0.59-0.84, 0.55-0.88, and 0.57-0.88, respectively. Among the algorithms, the logistic regression models performed best and had the highest AUC for six ADRs (range 0.67-0.83). The nausea-vomiting prediction models performed best with an AUC of 0.83 for the three algorithms. Conclusions: The prediction models for chemotherapy-induced ADRs were able to predict eight ADRs using EHR data. The logistic regression models were best suited to predict ADRs. The models developed in this study can be used to predict the risk of ADRs in patients receiving chemotherapy.