Background Diabetes affects more than 30 million patients across the United States. With such a large disease burden, even a small error in classification can be significant. Currently billing codes, assigned at the time of a medical encounter, are the “gold standard” reflecting the actual diseases present in an individual, and thus in aggregate reflect disease prevalence in the population. These codes are generated by highly trained coders and by health care providers but are not always accurate. Objective This work provides a scalable deep learning methodology to more accurately classify individuals with diabetes across multiple health care systems. Methods We leveraged a long short-term memory-dense neural network (LSTM-DNN) model to identify patients with or without diabetes using data from 5 acute care facilities with 187,187 patients and 275,407 encounters, incorporating data elements including laboratory test results, diagnostic/procedure codes, medications, demographic data, and admission information. Furthermore, a blinded physician panel reviewed discordant cases, providing an estimate of the total impact on the population. Results When predicting the documented diagnosis of diabetes, our model achieved an 84% F1 score, 96% area under the curve–receiver operating characteristic curve, and 91% average precision on a heterogeneous data set from 5 distinct health facilities. However, in 81% of cases where the model disagreed with the documented phenotype, a blinded physician panel agreed with the model. Taken together, this suggests that 4.3% of our studied population have either missing or improper diabetes diagnosis. Conclusions This study demonstrates that deep learning methods can improve clinical phenotyping even when patient data are noisy, sparse, and heterogeneous.