Doctor AI is a predictive model developed using recurrent neural networks (RNN) to forecast clinical events from electronic health record (EHR) data. The model was trained on longitudinal EHR data from 260,000 patients over 8 years, enabling it to predict diagnosis and medication categories for future visits. It also predicts the time until the next visit, aiding in clinical decision-making. The model uses multilabel prediction to identify multiple diagnoses and medications, achieving a recall@30 of 79%, significantly higher than several baselines. Doctor AI demonstrates strong generalizability, as it can be adapted from one institution to another without losing substantial accuracy. The model was tested on a dataset of primary care patients from Sutter Health, with data including encounter records, medication orders, and procedure codes. Medical codes were grouped into higher-level categories to reduce feature complexity. The model uses RNNs with gated recurrent units (GRUs) to learn patient representations and predict future events. It was trained on a large dataset and evaluated using metrics such as top-k recall and coefficient of determination (R²). The model outperformed baselines in predicting diagnosis and medication codes, as well as the time until the next visit. Doctor AI also showed effectiveness in transfer learning across different medical institutions, indicating its potential for use in health systems with limited patient data. The model's ability to learn from large datasets allows it to improve prediction accuracy in smaller populations. The study highlights the potential of RNNs in learning complex sequential patterns from EHR data, offering a practical tool for clinical decision support. However, challenges remain in accurately predicting time intervals due to the influence of various personal factors. Future work aims to enhance the model's performance to better assist physicians in clinical settings.Doctor AI is a predictive model developed using recurrent neural networks (RNN) to forecast clinical events from electronic health record (EHR) data. The model was trained on longitudinal EHR data from 260,000 patients over 8 years, enabling it to predict diagnosis and medication categories for future visits. It also predicts the time until the next visit, aiding in clinical decision-making. The model uses multilabel prediction to identify multiple diagnoses and medications, achieving a recall@30 of 79%, significantly higher than several baselines. Doctor AI demonstrates strong generalizability, as it can be adapted from one institution to another without losing substantial accuracy. The model was tested on a dataset of primary care patients from Sutter Health, with data including encounter records, medication orders, and procedure codes. Medical codes were grouped into higher-level categories to reduce feature complexity. The model uses RNNs with gated recurrent units (GRUs) to learn patient representations and predict future events. It was trained on a large dataset and evaluated using metrics such as top-k recall and coefficient of determination (R²). The model outperformed baselines in predicting diagnosis and medication codes, as well as the time until the next visit. Doctor AI also showed effectiveness in transfer learning across different medical institutions, indicating its potential for use in health systems with limited patient data. The model's ability to learn from large datasets allows it to improve prediction accuracy in smaller populations. The study highlights the potential of RNNs in learning complex sequential patterns from EHR data, offering a practical tool for clinical decision support. However, challenges remain in accurately predicting time intervals due to the influence of various personal factors. Future work aims to enhance the model's performance to better assist physicians in clinical settings.