There have been few studies assessing long-term outcomes of asthma based on regular follow-up data.We aimed to demonstrate clinical outcomes of asthma by multidimensional analyses of a long-term real-world database and a prediction model of severe asthma using machine learning.The database included 567 severe and 1,337 nonsevere adult asthmatics, who had been monitored during a follow-up of up to 10 years. We evaluated longitudinal changes in eosinophilic inflammation, lung function, and the annual number of asthma exacerbation (AE) using linear mixed effects model. Least absolute shrinkage and selection operator logistic regression was used to develop a prediction model for severe asthma. Model performance was evaluated and validated.Severe asthmatics had higher blood eosinophil (P = 0.020) and neutrophil (P < 0.001) counts at baseline than nonsevere asthmatics; blood eosinophil counts showed significantly slower declines in severe asthmatics than nonsevere asthmatics throughout the follow-up (P = 0.009). Severe asthmatics had a lower level of forced expiratory volume in 1 second (P < 0.001), which declined faster than nonsevere asthmatics (P = 0.033). Severe asthmatics showed a higher annual number of severe AE than nonsevere asthmatics. The prediction model for severe asthma consisted of 17 variables, including novel biomarkers.Severe asthma is a distinct phenotype of asthma with persistent eosinophilia, progressive lung function decline, and frequent severe AEs even on regular asthma medication. We suggest a useful prediction model of severe asthma for research and clinical purposes.