A paper published by Google on arXiv titled "Scalable and Accurate Deep Learning for Electronic Health Records" (Alvin Rajkomar et al.) presents an innovative approach to utilizing raw electronic health record (EHR) data. The study leverages the FHIR (Fast Healthcare Interoperability Resources) format, a standardized method for exchanging healthcare information, to build deep learning models that accurately predict various medical events.
The research highlights how traditional predictive modeling often requires extensive manual effort to extract relevant features from normalized EHR data, which can lead to loss of critical patient information. Instead, the authors propose using the raw EHR data in a structured sequence format based on FHIR. This approach enables accurate predictions across multiple medical centers without requiring centralized data coordination.
The model was tested using de-identified EHR data from two U.S. academic medical centers, involving 216,221 adult patients who stayed in the hospital for at least 24 hours. The dataset contained over 46 billion data points, including clinical instructions. The results were impressive: the deep learning models achieved high accuracy in predicting in-hospital mortality (AUROC 0.93–0.94), 30-day unplanned readmissions (AUROC 0.75–0.76), extended hospital stays (AUROC 0.85–0.86), and final diagnoses (frequency-weighted AUROC 0.90). In all cases, these models outperformed traditional statistical methods.
Additionally, the researchers introduced a neural network attribution system that helps clinicians understand how predictions are made, providing transparency in the decision-making process. They believe this approach can be widely applied in different clinical settings, with key evidence clearly highlighted in the patient's records.
As part of their work, the team emphasized the need for a protocol buffer tool to handle large-scale machine learning tasks using FHIR. This tool would allow efficient serialization of vast amounts of data, making it easier to analyze and process large datasets.
Google recently announced that the FHIR protocol buffer tool has been open-sourced. According to the blog post, over the past decade, healthcare data has transitioned from paper-based to digital formats. However, challenges remain, such as inconsistent data representations across vendors, varying coding practices for similar medical terms, and fragmented data spread across multiple tables.
FHIR aims to address these issues by providing a scalable and standardized framework for exchanging health data. While FHIR has become a widely adopted standard, large-scale machine learning requires additional tools. The newly open-sourced protocol buffer tool supports Java and will soon include support for C++, Go, and Python. It also includes features for configuration files and tools to convert legacy data into FHIR format, further enhancing its usability.
Pcb Pluggable Terminal Block Connector ,Pluggable Terminal Block,Contact Pluggable Terminal Blocks ,Pluggable Screw Terminal
Cixi Xinke Electronic Technology Co., Ltd. , https://www.cxxinke.com