**Doctor AI: Predicting Clinical Events via Recurrent Neural Networks** **Abstract:** Doctor AI is a generic predictive model that leverages large historical data in electronic health records (EHR) to predict observed medical conditions and medication uses. The model, developed using recurrent neural networks (RNN), was applied to longitudinal time-stamped EHR data from 260,000 patients over 8 years. Encounter records, including diagnosis codes, medication codes, and procedure codes, were input to the RNN to predict future diagnoses and medication categories for subsequent visits. Doctor AI assesses patient history to make multilabel predictions. Evaluations on a blind test set showed that Doctor AI can perform differential diagnosis with up to 79% recall@30, significantly outperforming several baselines. The model also demonstrated great generalizability, as it could be adapted from one institution to another without significant loss in accuracy. **Introduction:** The challenge in healthcare is the availability of massive amounts of patient data but limited time and tools for physicians. Intelligent clinical decision support systems aim to anticipate specific information at the point of care. EHR data, now commonplace in U.S. healthcare, represent longitudinal patient experiences and are increasingly used to predict future events. While existing predictive models focus on specialized outcomes, day-to-day clinical practice involves a heterogeneous mix of scenarios requiring different prediction models. Doctor AI addresses this by using RNNs to represent patient status and predict diagnoses, medication orders, and visit times. **Methods:** The RNN model for multilabel point processes was designed to learn effective patient representations. The model uses RNNs with Gated Recurrent Units (GRUs) to predict diagnosis and medication codes and the time duration until the next visit. The input to the RNN includes multi-hot vectors of medical codes and the duration since the last event. The model is trained using a joint loss function that combines cross entropy for code prediction and squared loss for time duration prediction. **Results:** Experiments on a large EHR dataset showed that Doctor AI achieved 79.58% recall@30, significantly outperforming baselines. The model's performance improved with longer patient records and more frequent visits. Knowledge transfer from one hospital to another was also demonstrated, showing that learned representations could be adapted to new datasets. **Conclusion:** Doctor AI is a powerful tool for predicting clinical events using RNNs. It mimics human doctors' predictive power and provides clinically meaningful diagnostic results. Future work will focus on improving the model's practical utility by addressing limitations such as the importance of incorrect predictions in medical practice.**Doctor AI: Predicting Clinical Events via Recurrent Neural Networks** **Abstract:** Doctor AI is a generic predictive model that leverages large historical data in electronic health records (EHR) to predict observed medical conditions and medication uses. The model, developed using recurrent neural networks (RNN), was applied to longitudinal time-stamped EHR data from 260,000 patients over 8 years. Encounter records, including diagnosis codes, medication codes, and procedure codes, were input to the RNN to predict future diagnoses and medication categories for subsequent visits. Doctor AI assesses patient history to make multilabel predictions. Evaluations on a blind test set showed that Doctor AI can perform differential diagnosis with up to 79% recall@30, significantly outperforming several baselines. The model also demonstrated great generalizability, as it could be adapted from one institution to another without significant loss in accuracy. **Introduction:** The challenge in healthcare is the availability of massive amounts of patient data but limited time and tools for physicians. Intelligent clinical decision support systems aim to anticipate specific information at the point of care. EHR data, now commonplace in U.S. healthcare, represent longitudinal patient experiences and are increasingly used to predict future events. While existing predictive models focus on specialized outcomes, day-to-day clinical practice involves a heterogeneous mix of scenarios requiring different prediction models. Doctor AI addresses this by using RNNs to represent patient status and predict diagnoses, medication orders, and visit times. **Methods:** The RNN model for multilabel point processes was designed to learn effective patient representations. The model uses RNNs with Gated Recurrent Units (GRUs) to predict diagnosis and medication codes and the time duration until the next visit. The input to the RNN includes multi-hot vectors of medical codes and the duration since the last event. The model is trained using a joint loss function that combines cross entropy for code prediction and squared loss for time duration prediction. **Results:** Experiments on a large EHR dataset showed that Doctor AI achieved 79.58% recall@30, significantly outperforming baselines. The model's performance improved with longer patient records and more frequent visits. Knowledge transfer from one hospital to another was also demonstrated, showing that learned representations could be adapted to new datasets. **Conclusion:** Doctor AI is a powerful tool for predicting clinical events using RNNs. It mimics human doctors' predictive power and provides clinically meaningful diagnostic results. Future work will focus on improving the model's practical utility by addressing limitations such as the importance of incorrect predictions in medical practice.