Taipei Medical University Clinical Research Database: a collaborative hospital EHR database aligned with international common data standards

Database development

The Clinical Data Centre (CDC) at the TMU Office of Data Science is a collaborative group made up of experts in data science, pharmacists and practising physicians. They have joined forces to create the research database. The TMUCRD database is filled with information collected during regular hospital care, meaning it does not cause any extra work for healthcare providers or disrupt their usual routines. The data have been gathered from various sources and linked, including:

• Archives from hospital information system (HIS) databases.

• Taiwan Cancer Registry database.

• Taiwan Death Registry database.

During the data collection period, information was gathered from three distinct HIS—TMUH, WFH and SSH. These systems served as the origin of clinical data, comprising various elements such as:

• Different types of forms like outpatient, inpatient and emergency records.

• Results of ordered measurements.

• Medications prescribed by clinicians/physicians.

• Details about procedures performed and associated fees.

• Patient demographic data including birthdates, zip codes, height, weight, blood pressure readings (systolic blood pressure, diastolic blood pressure), temperature for each hospital visit and in-hospital mortality.

• Recorded notes such as discharge summaries and reports from examinations such as radiology, cardiology and pathology.

• Medical images in Digital Imaging and Communications in Medicine (DICOM) format, which include X-rays, CT scans, MRIs and ultrasounds.

With the exception of data specifically collected for research purposes, the data were extracted and organised into database tables with structures distinct from those of the HISs. These data are stored individually for each hospital and are differentiated using a suffix denoting their source. For instance, TMUH’s outpatient visits are stored in the OPD_BASIC_T table while WFH’s and SHH’s outpatient visits are stored in the OPD_BASIC_W and OPD_BASIC_S tables, respectively. However, patient data can still be cross-referenced across hospitals using their pseudoidentification, represented by the ‘ID_NO’.

We acquired information about mortality occurring outside the hospital environment by referring to the Taiwan Death Registry database, which is maintained by the Taiwan Ministry of the Interior.21 Additionally, we have established a link between the TMUCRD and the Taiwan Cancer Registry, a dataset offered by the Taiwan Ministry of Health.22 This linkage allowed us to identify patients who were diagnosed with various forms of cancer and had visited any of the three hospitals in our study.

The TMUCRD vocabulary contains various terms, and the team at the CDC has worked to link these terms with standardised dictionaries within the database. As an example, the codes used for laboratory tests and medications in TMUCRD, which are recognised by Taiwan’s National Health Insurance (NHI), have been connected to codes in LOINC23 and RxNorm,24 respectively. These efforts have been made to adapt TMUCRD into widely accepted data formats, such as the Observational Medical Outcomes Partnership CDM (OMOP CDM). This adaptation enables the use of consistent tools and methodologies.25

Deidentification

Prior to being integrated into the TMUCRD database, the data underwent a deidentification process to adhere to the standards set by the Health Insurance Portability and Accountability Act (HIPAA). The initial step was conducted independently by the Centre for Management and Development (CMD) at TMU.26 This deidentification was achieved using structured data techniques.27 The process for structured data involved the elimination of eighteen specific data elements that could potentially identify individuals, as outlined in HIPAA. This removal included details such as patient names, phone numbers, addresses and dates. Notably, for the birth dates, only the year and month were retained for each patient, ensuring further privacy.

Moreover, an additional layer of deidentification was implemented by introducing randomisation to the variables within each data table. Essentially, we combined the initial pseudoidentification with a randomly generated salt-key, which consists of data from one or multiple variables associated with each patient. This salt-key serves as an additional input to a one-way function that hashed the pseudoidentification. Additionally, we employed checksum functions using MD5, SHA1 and SHA256 algorithms, which are types of hash functions. This process was completed before providing the data to each respective study principal investigator (PI). It is important to note that the components of this deidentification system are consistently expanded to accommodate new data as it is obtained.

The code used to create the TMUCRD introduction website and its accompanying documentation is accessible solely to individuals associated with TMU, including the PIs. The link to access this code is available.28

CDM conversion

The OMOP CDM serves as a standardised structure for organising observational medical data. Its purpose is to ensure the reliable analysis and utilisation of medical information for research purposes. This model includes standardised vocabularies that establish uniform terminology usage across different medical areas.29 Essentially, it provides a systematic framework for converting varied healthcare data into a shared format, facilitating consistent analysis across diverse data sources and research investigations.30 Starting in January 2021, the TMUCRD database embarked on a journey to adapt its data to the OMOP CDM standard. This transition was facilitated with the support of the Observational Health Data Sciences and Informatics (OHDSI) global initiative. The amalgamation of data from all three affiliated hospitals led to the naming of the database as the TMU-CMD.

Technical validation

To maintain the close representation of the original data collected from the three affiliated hospitals, we aimed to minimise significant changes to the structure of TMUCRD while achieving the necessary level of deidentification and data schema.

We adhered to the best practices in scientific computing whenever feasible. The development of TMUCRD was managed with version control, ensuring that changes were well tracked and documented. Issue tracking was implemented to transparently document any limitations in the data or code and address them appropriately. We actively encourage the research community to report and address any issues they come across. Furthermore, we have established a system for minor updates to the database.

The process of converting to TMU-CDM, which is the TMU-CDM, was carefully validated. This validation process followed the guidance of the OHDSI global initiative, particularly the SOS project.31 This rigorous approach ensured the accuracy and reliability of the conversion process.

Data records and tables

TMUCRD is a relational database that comprises 29 individual tables for each of the three hospitals involved. Additionally, there are two linked tables that connect with the Taiwan government (online supplemental appendix figure S1). Within each hospital, these tables are connected using identifiers, typically employing hospital-specific and visit-specific IDs (eg, CHR_NO, FEE_NO) for each patient. For example, ‘CHR_NO’ refers to a unique patient in a hospital, and ‘CHR_NO and FEE_NO’ refer to a unique outpatient visit or a unique admission to the hospital.

To ensure the accuracy of transformations and to maintain the fidelity of the original hospital data, we were cautious not to make assumptions about the underlying data. This approach enabled TMUCRD to faithfully represent the raw hospital data. The distribution of data across different categories and tables is outlined in table 1. Broadly, TMUCRD encompassed nine categories: demographics, diagnoses, medications, procedures, measurements, image examinations and radiology, surgeries, cancer-related data, pathology and vocabulary. The vocabulary category contained dictionary tables providing definitions for various identifiers. For instance, in the OPD_MED table, each row was associated with a unique MED_CODE representing a medication concept as outlined by Taiwan’s NHI regulations. By connecting the OPD_MED and MED_BASIC tables using MED_CODE, we could discern the concept behind a specific MED_CODE.

Furthermore, information pertaining to patient visits was stored in the diagnosis category through tables such as OPD_BASIC, IPD_BASIC and EPD_BASIC. Other categories contained data linked to patient treatments, medications, procedures, measurements, and image examinations and radiology. In some cases, it was possible to merge tables. For instance, the OPD_FEE and OPD_EXPER tables both contained details about measurements and could be combined. However, we chose to maintain the independent tables for clarity, given the substantial disparities in data sources and content.

Patient demographic characteristics

TMUCRD encompassed data from a 4.15 million unique patients who visited three hospitals in northern Taiwan from 2004 to 2021. This data compilation involved more than 61.5 million outpatient visits, around 3 million emergency visits, and roughly 1.1 million hospital admissions. The average age of patients during their initial visit had a mean of 38.8 years with an SD of 20, and a median of 37.2 years within an IQR range of 24.5–53.2 years. A majority of the patients were female, making up 53.9% of the total. The overall TMUCRD dataset showed a mortality rate of 7.5%, as linked with the National Death Registry, while the in-hospital mortality rate was 0.88%. On average, patients spent around 4 days in the hospital (eg, with a median of 4 days and an IQR range of 2–8 days). The mean observation period for patients was approximately 1451 days, with an SD of 1940 days. For a detailed breakdown of the patient population within each hospital, it can refer to table 2.

Diseases

Table 3 presents information concerning disease categories within each hospital. A total of 19 disease systems, categorised using the International Classification of Disease, 9th and 10th Revision, Clinical Modification (ICD-9-CM and ICD-10-CM), were subjected to descriptive analysis. Without considering the analysis of symptoms, signs, ill-defined conditions (ie, ICD9: 780–799; ICD10: P00–R99), and supplementary classifications encompassing factors influencing health status and healthcare interactions (ie, ICD9: V01–V91; ICD10: U00–U85, Z00–Z99); the three most prevalent disease categories, were as follows:

• Diseases of the digestive system (ie, ICD9: 520–579; ICD10: K00–K95), which constituted 26.7% of all patients.

• Diseases of the musculoskeletal system and connective tissue (ie, ICD9: 710–739; ICD10: M00–M99), accounting for 21.2% of all patients.

• Diseases of the respiratory system (ie, ICD9: 460–519; ICD10: J00–J99), representing 20.4% of all patients.

Medications

Medications were organised into different groups based on the specific organ or system they impact, categorised at various levels. In table 3B, we can observe the prevalence rates of 14 distinct medication groups, classified using the Anatomical Therapeutic Chemical classification system at the first level. The majority of medications utilised fell within class N (nervous system), constituting 36% of usage. Following closely were class M (musculoskeletal system), class A (alimentary tract and metabolism) and class J (anti-infective for systemic use), accounting for 35.1%, 31% and 27.4% of usage, respectively.

Laboratory types

Figure 2 presents data regarding the count of laboratory tests conducted in the outpatient department, categorised by years, for each individual hospital within the TMUCRD. Additionally, the total count of patients who underwent these tests is also displayed. The range of laboratory tests varied widely, spanning from 1.15 to 3.8 million over the course of 18 years at WFH. Notably, the number of tests notably rose in 2021, reaching 3.04 million and 2.21 million for SHH and TMUH, respectively. For a more comprehensive breakdown of this information, it can refer to online supplemental appendix, table S1.

Request permission

Data access

TMUCRD is available in a range of specially structured SAS files. These files typically reside within SAS libraries and can be processed using SAS software V.9.4 (SAS Institute). Additionally, the data from these files can be imported into database systems like MySQL, PostgreSQL or MSSQL Server. Given that the database holds intricate information about patients’ clinical care, it necessitates appropriate handling with due care and respect. Researchers aiming to access the data must follow a formal procedure, outlined on the TMU-CDC website.35 There are specific prerequisites and steps to be fulfilled before access is granted:

• TMUCRD is accessible to PIs and research scholars affiliated with TMU and its associated hospitals.

• Applicants need to acquire and complete the research database application and case report forms.36 Subsequently, they must seek approval from the ethics committee through the TMU-eJIRB system.37 PIs, investigators and analytical personnel must be mentioned in the IRB and are required to sign the data use agreement. This approval process typically takes a minimum of 2 weeks. Once the application is sanctioned, the TMU-CDC will inform PIs and scholars via email, providing instructions for accessing the dataset.

• We provided two services:

  • ‘Data to go’: Applicants can receive released data to conduct their analyses. However, it noted that the number of patients included is less than 1% of the total population (approximately 30 000 patients), and no reports, such as radiology, pathology and discharge summaries, are provided.

  • ‘Report to go’: Applicants can access various types of data and can analyse their study research in a designated ‘clean room’. In this setting, individuals are prohibited from bringing in any external devices or items, ensuring that only verified reports intended for publication can be carried out.

Example usage

Since the TMUCRD became available for application in January 2020, the Data Office has received and successfully managed 289 consultation cases. Among these, 68% of the consultations originated from our university and affiliated hospital. These cases covered diverse subject areas including medicine (49%), pharmacy (22%) and other fields. Furthermore, a total of 527 applications were submitted, out of which 398 were granted approval.

The TMUCRD has served as the foundation for a wide array of research endeavours. These studies have delved into various subjects, such as using machine learning techniques to predict outcomes for patients with cancer, investigating the clinical implications of diabetes, exploring advanced CKD and assessing adverse outcomes following major surgeries.38–40

Collaborative research

Traditionally, many researchers and scholars work in isolation with their own data. However, we are actively transitioning towards a more collaborative and iterative approach to research. This shift improves result cross-validation and self-checking, bolstering research reliability. Additionally, we support pharmaceutical companies in postmarket surveillance, contributing to product evaluations and healthcare quality advancement.

Notably, TMU is Taiwan’s exclusive official member of the OHDSI initiative. OHDSI is committed to enhancing clinical medical data’s value through big data analysis and AI methods. It promotes multiparty research collaboration across various domains and addresses complex issues. A vital part of OHDSI’s mission is creating standardised CDMs to streamline data systems worldwide, ensuring consistency and comparability.

Starting from September 2020, the TMU CDC embarked on the OHDSI-CDM grafting project. With guidance from OHDSI headquarters, we organised a series of online conferences providing transnational technical support for the implementation of TMU-OHDSI OMOP CDM grafting. To date, we have actively participated in three large-scale multinational collaborative research projects. These projects encompass prognostic analysis of antihypertensive drug therapy, assessment of the effectiveness of anticoagulant drugs and the evaluation of cancer safety associated with H2 receptor antagonists.41

Conclusion

Data are the cornerstone of scientific research, profoundly affecting research outcomes. The TMU CDC is unwavering in its dedication to enhancing data quality, valuing larger data volumes, longer data periods, authenticity, diversity, integrity, standardisation, accessibility, privacy and robust data governance.

We continually develop advanced data management systems, including data processing tools, cloud-based decision interfaces, data sampling options and cross-institutional data quality checks. Our focus is on creating specialised databases, integrating diverse healthcare data like inspection reports, patient records, nutrition assessments and more. These databases adhere to international OMOP-CDM standards, supporting research on topics such as COVID-19, dementia, stroke, lung cancer, diabetes and CKD, exemplifying our commitment to diverse and impactful healthcare research.