Abstract
BACKGROUND: Nonsteroidal anti-inflammatory drugs (NSAID) are currently among the most prescribed medications worldwide to relieve pain and reduce inflammation, especially in patients suffering osteoarthritis (OA). However, NSAIDs are known to have adverse effects on the gastrointestinal system. If a gastric ulcer occurs, planned OA treatment needs to be changed, incurring additional treatment costs and causing discomfort for both patients and clinicians. Therefore, it is necessary to create a gastric ulcer prediction model that can reflect the detailed health status of each individual and to use it when making treatment plans. METHODS: Using sample cohort data from 2008 to 2013 from the National Health Insurance Service in South Korea, we developed a prediction model for NSAID-induced gastric ulcers using machine-learning algorithms and investigated new risk factors associated with medication and comorbidities. RESULTS: The population of the study consisted of 30,808 patients with OA who were treated with NSAIDs between 2008 and 2013. After a 2-year follow-up, these patients were divided into two groups: without gastric ulcer (n=29,579) and with gastric ulcer (n=1,229). Five machine-learning algorithms were used to develop the prediction model, and a gradient boosting machine (GBM) was selected as the model with the best performance (area under the curve, 0.896; 95% confidence interval, 0.883-0.909). The GBM identified 5 medications (loxoprofen, aceclofenac, talniflumate, meloxicam, and dexibuprofen) and 2 comorbidities (acute upper respiratory tract infection [AURI] and gastroesophageal reflux disease) as important features. AURI did not have a dose-response relationship, so it could not be interpreted as a significant risk factor even though it was initially detected as an important feature and improved the prediction performance. CONCLUSIONS: We obtained a prediction model for NSAID-induced gastric ulcers using the GBM method. Since personal prescription period and the severity of comorbidities were considered numerically, individual patients' risk could be well reflected. The prediction model showed high performance and interpretability, so it is meaningful to both clinicians and NSAID users.