The paper presents a scalable and accurate approach to predictive modeling using electronic health record (EHR) data, leveraging deep learning methods. The authors propose a representation of patients' entire raw EHR records based on the Fast Healthcare Interoperability Resources (FHIR) format, which allows for the extraction of a vast amount of information without the need for manual data harmonization. They demonstrate that deep learning models using this representation can accurately predict multiple medical events from multiple centers without site-specific data harmonization. The approach was validated using de-identified EHR data from two US academic medical centers with 216,221 adult patients. The models achieved high accuracy for tasks such as predicting in-hospital mortality, 30-day unplanned readmissions, prolonged length of stay, and all final discharge diagnoses. These models outperformed traditional, clinically-used predictive models in all cases. The study highlights the potential of deep learning to create accurate and scalable predictions for various clinical scenarios, including a case study where neural networks were used to identify relevant information from patient charts.The paper presents a scalable and accurate approach to predictive modeling using electronic health record (EHR) data, leveraging deep learning methods. The authors propose a representation of patients' entire raw EHR records based on the Fast Healthcare Interoperability Resources (FHIR) format, which allows for the extraction of a vast amount of information without the need for manual data harmonization. They demonstrate that deep learning models using this representation can accurately predict multiple medical events from multiple centers without site-specific data harmonization. The approach was validated using de-identified EHR data from two US academic medical centers with 216,221 adult patients. The models achieved high accuracy for tasks such as predicting in-hospital mortality, 30-day unplanned readmissions, prolonged length of stay, and all final discharge diagnoses. These models outperformed traditional, clinically-used predictive models in all cases. The study highlights the potential of deep learning to create accurate and scalable predictions for various clinical scenarios, including a case study where neural networks were used to identify relevant information from patient charts.