Emerging technologies, such as big data, the Internet of Things, cloud computing, and artificial intelligence, are profoundly changing medical and health care service models. Health and medical data sets, which are typically large with rapid growth, diverse data structure, multidimensional value density, high requirements of data credibility, and high concerns over data security, are often called health care big data. The term big data is used to represent not only a technology but also a resource and a method—a big data platform is a system and a tool that integrates data, tools, apps, and service.

Driven by national policies, high-income countries in Europe, such as the United Kingdom [1], and in North America, such as the United States [2], took the lead in building big data platforms for health care [3]. For example, Britain’s comprehensive platform integrates and applies data from 12 categories, including health, medical care, transportation, and environment, to support government decision-making [3]. The Big Data Analytics Platform built by the Czech Republic meets the data analysis requirements of its national public health service [4]. In recent years, China has built a national-level health information platform that can be connected to provincial-level health information platforms for data integration [5].

In addition to national-level big data platforms in health care that are driven by policies, large health care institutions have also built big data platforms to suit their own management needs. At present, a medium-sized health care institution in China generates 1 to 20 TB of health care data, and a large health care institution generates 300 TB to 1 PB of health care data every year. Some hospitals have used big data technology to build hospital-level scientific research platforms—for example, Ninewells Hospital and Medical School in the United Kingdom developed a research data management platform [6], Asan Medical Center in South Korea developed a clinical trial management system based on integrated data [7], and the People's Hospital of Peking University in China developed a hospital-wide big data platform for clinical research [8]—or to build hospital-wide data integration platform—for example, the Third Hospital of Peking University [9] and the Second Affiliated Hospital of Guangzhou University of Traditional Chinese Medicine [10]. Most hospitals use big data technology in analytics platforms for diseases (eg, for nasopharyngeal carcinoma [11]), gastrointestinal conditions [12], cancer [13], and cardiomyopathies [14].

Meta-analyses [15-18] have found that big data platforms in health care have had great impacts on have had great impacts on medical technology, medical service quality, and medical costs, but the actual construction process is not easy, with challenges arising in data structurization, security, standardization, storage, processing, and management [15,16]. As a result, the value mined from this type of data is currently limited [17], and in China, health care data utilization is still in the early stage [18], which necessitates the improvement of health care big data governance.

Health care data generated in large hospitals include (1) electronic medical records of clinical diagnosis and treatment, which contain diagnoses, prescriptions, surgical treatments, and examination findings; (2) data derived from health management or clinical research activities, including follow-up information, gene sequencing data, and physical examination data; (3) data related to hospital management, including patient wait times, bed turnover rate, medical equipment utilization efficiency, and revenue; and (4) data derived from web-based diagnosis and treatment services [19]. To make better use of medical data, the first task is to build a big data platform for data collection and governance to generate high-quality data assets. This will enable in-depth data analysis and mining, as well as the formation of rules of knowledge, allowing big data methods to benefit clinical practice, scientific research, and hospital management.

In addition to the technical considerations for platform construction, data management in large hospitals must take into account by data service management. Special attention should be paid to patient privacy protection and ethical concerns on data use. A hospital data platform is generally built for a specific purpose, such as scientific research, operations, data integration, or analysis, but large hospitals often require a comprehensive data platform—one that meets needs for medical treatment, education, scientific research, and management.

The overall objective of this study was to develop and implement a health care big data platform for a large hospital in China, with data governance as the core concept, to solve difficulties related to integration, calculation, storage, standardization, and security for multisource heterogeneous medical data. This platform will integrate the data from all operation systems in a hospital to generate high-quality data assets and form a massive high-dimensional medical database that can comprehensively support the clinical activities, scientific research, and management of the hospital.

From a technical aspect, construction of a big data platform needs to solve the following problems that are unique to the health care industry: (1) integration of multisource heterogeneous data from multiple independent information systems within the hospital; (2) computing power requirements brought by the development of machine learning and deep learning during data application; (3) difficulties in analyzing and utilizing information data (due to nonuniform data standards, inconsistent use of a master patient index, and the fact that most electronic medical records in China are written in natural language, it is impossible to directly analyze and use existing information data, and as a result, semantic interoperability must be improved with the use of medical terminology) and (4) data security and patient privacy protection.

The West China Hospital of Sichuan University (WCH) built a health care big data platform, which is referred to as the WCH-BDP or the platform hereafter. WCH is a world-renowned large hospital with more than 4800 beds and approximately 15,000 outpatients on average per day. WCH’s electronic medical record system was built in 2007; therefore, it has been in use for 14 years. The hospital has more than 100 departments and sections, and their clinical activities generate a massive amount of data.

Traditional information technology can no longer handle massive amounts of data that are continuously growing, causing difficulties in effectively integrating multisource heterogeneous data, the serious problem of isolated data islands, bottlenecks in data storage and calculation, difficulties in structurally utilizing medical records written in Chinese semantics, and high technical barriers in mining data from images, videos, and files. In 2017, WCH launched a hospital-wide health care big data platform project to address these difficulties.

The project focuses on the design and development of the platform architecture and does not involve any study of clinical data, so ethical declarations are not applicable.

The first step of platform construction was organizing a project management team. To ensure the performance of the platform, WCH set up 2 working groups—one for platform construction, and the other for platform management. The construction working group included the chief platform architect, information technology experts, system engineers, and data engineers. The hospital information center was responsible for building the platform. The construction working group focused on the following: (1) the application objectives of the platform, (2) data integration method and scope, (3) master patient index strategy, (4) reference standards for medical terms, (5) scope of the master data, (6) data model structure, and (7) system implementation and training. Project meetings were held weekly and status meetings were held monthly to summarize the progress of the phases until the platform was launched. The discussions in these meetings provided a solid foundation for platform design and development. After the launch of the platform, the construction working group also conducted regular training on system functions, maintenance, and the help manual.

The management working group included all stakeholders, such as the chief information officer, hospital management users, clinical users, and scientific research users. The management group was the data governance committee responsible for organizing and overseeing all data-related work, including data definitions, data benchmarking, data quality control, and data security, after the launch of the platform. The committee was in charge of formulating related management systems, work collaboration mechanisms, and procedure standardization.

The convergence of multisource heterogeneous data from all operation systems in the hospital was necessary to ultimately provide data to different medical services (Figure 1). Therefore, the design strategy was to use computing and storage devices with enough capacity to integrate data into the corresponding physical resources, with separate storage and computing processes based on the characteristics of the modality (eg, clinical data, image data, or genomic data). A data repository was built using data governance methods to meet the needs of different subject fields. The platform had to be able to combine all data service supports, such as data security service, terminology service for data governance, search engine service, virtualization service, and artificial intelligence service—all of which rely on the computing power and storage capacity of physical devices and data resources. Solutions for data integration and governance were core aspects of the construction process. Data governance, which is the core of data management and the basis for standardizing disordered data into highly available data, includes master index data governance, master data governance, metadata governance, data security management, and data quality control management [10,20,21].

Massive data were collectively stored in the storage devices of the WCH-BDP. To better support data analysis, storage and calculation environments were designed and provided according to the characteristics of each data modality. Generally, the amount of data structuration is negatively correlated with storage space and computing power (ie, the greater the data structuration, the lower the requirements for storage space and computing power) (Figure 2). Hence, data that are poorly structured require more storage space and stronger computing power.

Highly structured data do not require much storage space. They can be stored on a distributed storage device, and computations can be effectively run by conventional central processing units.

Semistructured data, such as the records of a patient’s major complaint, disease history, examination findings, and examination conclusions in Chinese electronic medical records do not occupy much storage space and can be stored in distributed storage devices. However, these data require natural language processing for analysis; therefore, both general graphics processing units (GPUs) and central processing units are needed to provide sufficient computing power.

Unstructured data that are large in total volume but small in individual volume (eg, the original image data in DICOM format generated by thin-layer scanning examinations such as radiology and ultrasound) require a relatively high storage space and must be stored in centralized storage devices (mostly network-attached storage) to save resources. To analyze and mine these data, characteristic modeling is performed mainly with machine learning or deep learning technologies, which requires the support of a large amount of GPU power and private GPU resources.

Unstructured data with a large individual volume, such as genomic data, are generally stored directly in object-based storage. Genomic sequencing generates not only a large amount of original sequencing data, but also, a large amount of data related to multiple processes in research on biological information. As a result, genomic data sets often have a tremendous size, and the storage space that is required is mostly measured in petabytes. Analysis and mining of the types of data stored in object-based storage require a lot of clustered computing power supplied by multiple high-performance GPUs, which necessitates the deployment of accelerated supercomputing power supported by GPU clusters.

Traditional big data platforms are planned and implemented in a unified way with storage and computing integration. The design of the WCH-BDP is the realization of storage and computation separation. The WCH-BDP first solves the problem of effective storage and management of massive genomic data files. Second, when researchers use genetic data for analysis, the scheduling software provided by the WCH-BDP loads the analysis data into the supercomputer environment for analysis. After the analysis is completed, the analysis results are retained, and the temporary storage space occupied by the analysis process is released.

In China, medical institution operations information database systems mainly include Microsoft SQL Server, Oracle database, MySQL database, Caché database, and MongoDB; other unstructured data are mostly stored in the form of files. The WCH-BDP achieves data integration through both real-time and non–real-time data entries. For real-time data integration, a master–slave database is generated, and the log data of the slave database are parsed and captured in real time. Real-time data integration is suitable for an operations information system with mostly structured data. Parsing the information in the slave database does not consume much physical resources and therefore does not affect the performance of the master database, which ensures the security and stability of data integration. Image data integration and entry can be completed by directly reading image files using the DICOM protocol. Genomic data and other data in the form of files are not integrated in real time and are entered via file transfer protocol.

The enterprise master patient index is the unique tag of a patient in a health care institution. One patient might have different enterprise master patient indexes in the different operating systems of a hospital because these systems were constructed independently. This necessitated a master patient index governance to standardize the enterprise master patient indexes. The governance of the master index data is achieved through the master index governance system. The governance system included 4 stages: data preparation, standardization strategy, data processing, and tracking or feedback (Figure 3).

The data governance strategy of the WCH-BDP mainly focused on 3 key values: ID number, name, and telephone number. The configurations and data processing of these 3 key values are shown in the following table (Table 1).

Master index data processing cannot rely only on the system for automatic completion. For example, in the fifth row of Table 1, both the Name and Telephone number are Equal, but the ID card number is Unequal, which might be because the same patient used a different ID card for registration. In this case, the last step for standardization of the master index is critical, during which the uniqueness of each master index can be ensured by manual adjustment after log analysis or by referring to the actual operational process.

Master data include all related data items listed in the data dictionary, such as health care institution code, drug code, diagnosis code, and anesthesia method [22] (Table 2). Master data governance is the aim to map and handle master data discrepancies caused by different system standards. Classification and reference standards for the master data to be processed are determined, and they are used to map relationships between data in the database. The results are published to users for subscription and application.

The WCH-BDP is an environment that can effectively support data services with voluminous data repositories and strong physical resources for data computation and storage. Under an overall ecology of data security management, the platform can provide support to data services such as computing, storage, and virtualization. Search engine services, terminology services, and artificial intelligence services can be effectively integrated to meet the needs of all health care operations.

The WCH-BDP was formally launched in November 2020, after 3 years of construction, for demand surveys, platform design, module development, and pilot operation. To date, the platform has formed a massive high-dimensional database with more than 12.49 million patients, 75.67 million visits, and 8475 data variables. The platform has brought hospital informatization to a new level and greatly improves the hospital’s overall big data capabilities.

The WCH-BDP has more than 20 PB data storage capacity and runs on a server cluster consisting of 252 physical servers: 117 servers are dedicated to computing and analyzing, and 38 are equipped with more than one GPU card. In the server cluster dedicated to computing, there are 320 CPU cores and 149 GPU cards. In total, the platform has computing powers that exceed 300 TFLOPS for clinic data, 600 TFLOPS for image data, and 1000 TFLOPS for genomic data.

The WCH-BDP integrates diagnosis and treatment data from clinical information systems, clinical research data from science and research information systems, and management activity data from management information systems. Of 134 hospital systems, the platform has integrated the data of 27 systems, including hospital information systems, laboratory information systems, radiology, ultrasound, web-based medical appointments, human resources, equipment and supply management, Internet medical service since January 2008. In addition, all image data in the picture and communication systems and more than 20,000 patients’ genomic data have been entered into the platform.

Using Fast Healthcare Interoperability Resources Reference Information Model 3.0 standards, data integrated into the WCH-BDP were reorganized and arranged based on the type of operational activities while considering the characteristics of China’s medical system. Finally, a database that includes 134 data charts for 18 subject fields was developed. The top 5 subject fields that were responsible for the most data in the database are listed in Table 6. To date, the database includes 8475 data variables and 6.272 billion rows of data records.

The platform has constructed a data asset directory for users by further standardizing the data through the governance of master data, metadata, terminology, and natural language processing. To ease the process of inquiry, the directory has a hierarchical tree structure that allows users to select the data items that they need using fuzzy searches. To minimize comprehension difficulties, the directory uses a data variable–naming scheme consistent with the names in the interface of the corresponding operations systems as much as possible; in addition, users can see the information of the data source, content, and range of values.

The data asset directory covers 13 fields and 1488 data variables (Table 7). Of the 13 fields, the original data variables in 9 fields are structured variables, and those in 3 fields (medical records, imaging examinations, and nursing care records) are semistructured text. Semistructured data were converted into poststructured derived variables by using approaches such as word segmentation, entity extraction, and semantic identification. For patient tag information, the app-oriented poststructured derived variables were generated using a data-mining algorithm.