Hospital readmission refers to the situation where a patient is re-hospitalized with the same primary diagnosis after discharge. Hospital readmission causes $26 billion preventable expense to the U.S. health systems annually and may indicate suboptimal care for patients. Accurately predicting patients' risk of readmission could alleviate those severe financial and medical consequences. Nevertheless, such prediction is challenging due to the dynamic and complex nature of the hospitalization process. The state-of-the-art studies apply statistical models with unified parameters for all patients and utilize static predictors in a study period, failing to consider patients' heterogeneous hazard and dynamic illness process. Fortunately, recent advances in deep learning enable personalized analysis on sequential data during a process. We present the first readmission predictive study with an innovative trajectory-aware deep learning framework to capture patients' varied hazard during a hospitalization process. We perform our analyses on a unique five-year national Medicare claims dataset including 3.6 million patients per year. Empirical analyses show that our model significantly outperforms the baseline models, with a precision of 77.08%, a recall of 99.24%, and an F1-score of 86.70%. This study contributes to IS literature and methodology by formulating the readmission prediction problem and developing a novel personalized readmission risk prediction framework, including a trajectory-aware representation and a trajectory-aware long short-term memory unit. This framework provides direct implications for health providers to assess patients' readmission risk and take early interventions to avoid potential negative consequences.