AbstractExisting data-driven bridge deterioration prediction methods mostly learn from abstract inventory data from a single source to predict the future conditions of bridges. Bridge inventory data [e.g., the National Bridge Inventory (NBI) data] are undoubtedly important but are not enough. They mainly describe bridge conditions using abstract, aggregated condition ratings that do not contain detailed information about bridge deficiencies and maintenance actions, thus limiting the performance of data-driven deterioration prediction and its usefulness in supporting bridge maintenance decision making. Learning from the wealth of heterogeneous (i.e., structured and unstructured) bridge data from multiple sources opens an unprecedented opportunity for enhanced data-driven bridge deterioration prediction. Such data include structured NBI and National Bridge Elements (NBE) data, structured traffic and weather data, and unstructured textual bridge inspection reports. To capitalize on this opportunity, this paper proposes a novel bridge data analytics framework, which allows for the extraction, integration, and analysis of structured and unstructured bridge data from different sources. At the cornerstone of this framework is a proposed deep learning–based bridge deterioration prediction method for analyzing and learning from the integrated bridge data to predict bridge deterioration. The proposed method includes three primary components: manifold learning for embedding the integrated bridge data into a low-dimensional dense space, cost-sensitive learning for modulating the misclassification costs to address the class imbalance in the data, and recurrent neural networks for learning from the embedded and balanced data from past years to predict the conditions of the primary bridge components (decks, superstructures, and substructures) in the next year. The method was evaluated in predicting the condition ratings of the decks, superstructures, and substructures of 2,646 bridges in the state of Washington. It achieved an average macroprecision and macrorecall of 89.9% and 85.8%, which are 15.0% and 22.4% higher than those achieved by learning from only NBI data.