Artificial intelligence (AI) is increasingly being used to turn enormous amounts of healthcare data into insights that support identification of patients at high risk for adverse events. This study presents an application of AI using clinical claims data to evaluate the performance of a model predicting falls, a highly prevalent and costly healthcare event. The study sample included 500,000 Medicare Advantage beneficiaries, sampled from a nationally representative claims database, who had at least 10 months continuous enrollment in 2015-2017. A deep neural network model was constructed to take 2-year-long sequences of patient diagnoses, medications and procedures as input, and output the probability of a fall in the third year. A Time-Aware Hierarchical Attention Model was designed using inception convolutional neural network with Attention at claim-level, and Time-Aware Long-Short-Term Memory with Attention at encounter-level. This model takes into account the relationship between long-term medical history and current status, and also the time intervals between medical events (rather than treating events as evenly spaced). Model performance was assessed by area under receiver operating characteristic curve (AUC), precision-recall curve and reliability diagram. The model resulted in an AUC of 0.791. Under a 0.2 threshold in the precision-recall curve, the model resulted in 78% recall with 41% precision. A 0.7 threshold yielded 26% recall and 70%precision. The model is also well calibrated with Brier score as 0.049. Claims associated with a significant increase in falls risk included limb thrombus, osteoarthritis, urinary disorder with frequent night urine, and chronic obstructive pulmonary disease. Deep neural network modeling can be used to leverage large amounts of data and identify increased risk of falls. It can help decision makers target appropriate interventions and medical resources to reduce the incidence of adverse events and reduce the concomitant medical costs.