According to the Lancet Commission on Global Surgery, 4.2 million people worldwide die within 30 days of surgery per year. This major mortality risk exceeds that of human immunodeficiency virus, malaria, and tuberculosis combined [1,2]. Many of these deaths are potentially preventable [3–5]. This is in addition to other nonfatal perioperative complications, which can cause negative long-term health outcomes [6,7] and contribute to high healthcare costs [8,9]. Hence, improving perioperative outcomes is essential, as not only individual patients but also health systems as a whole benefit.

Both research and quality improvement in healthcare require an accurate understanding of patient profiles and care trajectories. To enable such efforts, a curated registry comprising high-quality real-world data for the entire perioperative period, including the preoperative, intraoperative, and postoperative periods of hospitalization, is paramount. Such a registry would provide the essential data for retrospective studies of factors influencing patient outcomes and prospective evaluations of clinical interventions. By consolidating and organizing perioperative data in an integrated database, researchers could easily access and analyze a large amount of collected data, which could lead to more accurate predictive models, an improvement in patient outcomes, and ultimately, improvements in clinical decision-making. In addition, an integrated perioperative registry would ensure access to data that is standardized and high quality, through the use of quality assessment methods. This is critical for accurate data analysis and the development of machine-learning models. By enforcing data standards and quality checks, researchers can minimize the risk of bias and errors in statistical analysis and develop machine-learning models. Furthermore, the use of a registry could address the data sharing and privacy concerns associated with the use of large institutional datasets. A secure and centralized location for data storage ensures that patient data are properly de-identified and protected and can only be accessed by authorized personnel.

Currently available high-resolution datasets include those derived from intensive care unit (ICU) stays, such as the frequently cited Medical Information Mart for Intensive Care (MIMIC) III and IV datasets and the eICU, Amsterdam UMCdb, and HiRID databases [10–12]. Other datasets, such as the Multicenter Perioperative Outcomes Group (MPOG) dataset, have focused on perioperative information in the western populations [13,14]. These datasets have enabled important research such as predictive analytics using machine learning and waveform processing [15]. However, no large high-resolution databases integrating all patient-care areas throughout the perioperative period are currently available. To address this gap, we established an integrated, standardized, and curated perioperative registry database at our institution.

Singapore General Hospital (SGH) is a quaternary care and academic medical center. It is the largest hospital in Singapore, with approximately 28,000–30,000 surgeries conducted each year. A comprehensive electronic medical record (EMR) has been implemented at SGH along with a fully digital anesthesia information and monitoring system (AIMS) for the operating room. Every patient’s surgical journey generates a digital footprint that encompasses the medical doctors, nursing, and allied health consultation notes; laboratory and imaging results; physiological patient monitoring; medication or blood product prescriptions; operation time; and length of hospital stay. These data exist in the following formats: structured data (e.g., patient demographics and comorbidities), unstructured data (e.g., patient communication notes), high-resolution physiological time-series data (e.g., intraoperative heart rate and blood pressure), and imaging data (e.g., radiology images).

This study aims to describe the methodology used to set up the perioperative data mart used at SGH, the Perioperative and Anesthesia Subject Area Registry (PASAR), which combines data from the preoperative, intraoperative, and postoperative periods, allowing for a seamless investigation of a patient’s entire journey from the operating room to the ICU. An overview of the case mix and available data within this registry is also presented.  

The PASAR covers all patients who undergo surgery at SGH. It was initiated in 2016 and continues to be populated with data to the present day.

Approval for the database was granted by the SingHealth Centralised Institutional Review Board (Singhealth CIRB 2014/651/D, 2020/2915, and 2021/2547). The requirement for individual patient consent was waived as the data are collected during routine clinical care and de-identified. The PASAR operates in compliance with Singapore’s Personal Data Protection Act and the Human Biomedical Research Act as well as other ethical guidelines based on the Declaration of Helsinki, 2013 [16–18]. The pipeline from routine clinical data and de-identification protocols are described below.

The main data source for the registry is the EMR (Sunrise Clinical Manager, Allscripts), AIMS (WinChart^®, WinChart Health Informatics), and administrative billing (SAP ISH, SAP SE) systems at SGH. All data from internal sources are acquired during routine patient care (i.e., no specific data are collected for research). The Singapore National Registration Identification number is a national identifier unique to each individual and is the common identifier for data contained in various systems throughout Singapore.

All hospital data are stored in raw format in our enterprise data warehouse system (SingHealth-IHiS Electronic Health Intelligence System — eHINTS). This structured query language (SQL) database allows for data staging before it is uploaded to the registry. Manual data entries are unavailable. All data are stored securely within an approved access-controlled facility at SGH in a manner compliant with local privacy and data protection laws. The database is synchronized with the National Registry of Births and Deaths to ensure that all out-of-hospital deaths are captured.

All patients undergoing surgery at SGH require a preoperative anesthesia assessment. A standardized electronic clinical document has been used since 2016 to identify individuals who qualify for inclusion in the database. All patients aged ≥ 18 years who undergo surgery under anesthesia care at SGH are included.

De-identification processes are performed by a trusted third party within SGH that is not involved in research-related activities using PASAR data. This trusted third party acts in accordance with Singapore’s Personal Data Protection Act and Human Biomedical Research Act [16,17].

Our institution mandates the de-identification of 15 direct identifiers, which include case and visit numbers, death dates, postal codes, and names in structured fields and free-text data. De-identification of case and visit numbers is performed by pseudonymization using a hashing algorithm (SHA-256) that returns substitute values. Before hashing, the identifiers are salted with project-specific values at the beginning and end. This process of salting and hashing the case and visit numbers is used consistently to ensure that the study team is able to link cases across all tables using substitute values. Death dates are de-identified by generalization to months and years (i.e., the exact day is censored).

In addition to fields that explicitly contain these 15 direct identifiers, other high-risk data fields, such as unstructured text (which can contain names and other identifiers) are completely concealed.

For each individual, data are collected during the preoperative, intraoperative, and postoperative periods.

The clinical variables included in the data registry have expanded over time. At the time the PASAR was initiated in 2016, data on preoperative medical history and laboratory results as well as important clinical outcomes such as length of hospital stay, ICU admissions, and death were included [19,20]. Currently, additional data routinely captured during preoperative anesthesia assessment visits are included, such as patient demographics, comorbidities, preoperative laboratory test results, and surgery details; intraoperative AIMS data including all administered medications, procedures performed, and high-velocity vital signs time series data; and postoperative data including post-anesthesia care unit, ICU, 30-day vital signs, and fluid balance data. Furthermore, discharge summaries in both free-text entries and structured ICD-10 codes are included for all case entries. A full data dictionary is available for accredited users.

Establishing clear data ownership ensures that the data are properly collected, stored, and maintained, and that all stakeholders are aware of their roles and responsibilities regarding the data. Ownership of the PASAR lies with the first author (HRA), who is responsible for ensuring proper delegation to authorized nominees and users. Only authorized users with a legitimate need to access the data are granted access, and access is monitored and audited regularly to ensure that the data is not being accessed inappropriately. The Data Science and Artificial Intelligence Laboratory at SGH is a data management team that is responsible for establishing data quality checks and ensuring that the data are stored securely.

Data sharing is essential to facilitate research collaboration and advance perioperative research. However, data sharing must be performed in a manner that protects patient privacy and confidentiality. This includes establishing data-sharing agreements that specify how data can be used and who can access it. All data are anonymized before being released to researchers, and researchers must sign a data-use agreement that prohibits re-identification and unauthorized redistribution of the data. The authors can be contacted for details on the data-sharing agreement and protocols.  

As of December 2022, the PASAR consists of a relational database containing 26 tables linked by the patient’s case number (a unique identifier for every admission), operation session identifier, and operation identifier (as multiple operations can be performed in a single session). The database is segmented into three schemata based on the perioperative period the data were collected: pre_op (preoperative), intra_op (intraoperative), and post_op (postoperative). The pre_op, intra_op, and post_op schemata consist of 5, 7, and 14 tables, respectively. Detailed entity relationship diagrams of these three schemata and variable descriptions are available in Supplementary Fig. 1 and Supplementary Table 1. A total of 153,312 patients were admitted, among which 168,977 operation sessions containing 183,687 operations were conducted. Fig. 1 shows these data broken down by year.

The profiles of the patients included in the study are listed in Table 1. Discrete variables are reported as counts and percentages and continuous variables are reported as median (Q1, Q3). These measures provide a concise overview of patient characteristics, allowing for other studies or populations to be easily compared. This patient cohort consisted of 153,312 patients, with a median age of 60.0 years (45.0, 69.0) and a balanced sex distribution (females: 50.2%). The majority of patients were Chinese (72.4%), followed by Indian (10.6%) and Malay (9.6%). Almost half of the patients had hypertension (44.9%) and a significant number of cardiovascular comorbidities were reported, such as ischemic heart disease (12.7%), kidney dysfunction (5.8%), diabetes mellitus on insulin (4.8%), and congestive heart failure (3.3%). Operation characteristics are provided in Table 2. Summary statistics were calculated on a per-operation basis. Counts and percentages are presented for discrete variables, while median (Q1, Q3) are presented for continuous variables. Most patients had American Society of Anesthesiologists physical status scores of 2 (50.0%) and 3 (35.1%), while 14.6% of the patients underwent emergency surgeries.

Table 3 provides information on surgical outcomes. Summary statistics were calculated on a per-operation basis. Counts and percentages are presented for discrete variables, while median (Q1, Q3) are presented for continuous variables. Clinically relevant outcomes included the operation time, ICU length of stay, and 30 days and 1-year mortality. The median ICU length of stay after emergency surgery was 3 days longer than that after elective operations (5.1 vs 1.9 days); however, the median operation time was shorter (0.9 vs 1.2 h).