This was a retrospective observational, single-centre study. This study was approved by Singapore Health Services’ Centralized Institutional Review Board.
Study setting and population
Singapore has a mixed healthcare delivery system consisting of both public and private providers, with the public sector handling over 80% of the population’s healthcare needs . There were a total of 5 adult care public hospital EDs in Singapore in 2008. Subsequently, two additional EDs started operations in 2010 and 2015 respectively with the opening of two new public hospitals, bringing the total number of EDs to 7 in 2017.
The public healthcare system in Singapore consists of three integrated clusters. The study hospital, Singapore General Hospital (SGH), is part of the Singapore Health Services (SingHealth) cluster, which delivers healthcare predominantly for the population in the Eastern region. We performed the study using a database from the SGH, the largest and oldest tertiary medical centre in Singapore, with comprehensive clinical services and over 1700 inpatient beds. Annually, the SGH ED receives more than 120,000 ED visits, over 40,000 of which are converted to inpatient admissions. As a percentage of the total emergency department visits of all public hospitals in Singapore, SGH ED visits represented 17% in 2008 and 13% in 2017. The analysis was based on extracted data from SGH’s electronic medical health system, namely Singhealth Electronic Health Intelligence System (eHints). Detailed information from other public or private hospitals were not available. The data was recorded as per individual emergency admission episodes. Multiple emergency admission episodes from the same patient are considered separate individual episodes. All patients who underwent emergency admissions at SGH from 1 January 2008 to 31 December 2017 were included in this study. Patients at SGH below the age of 18 were excluded.
Selected variables included two demographic variables, three administrative variables, and 18 clinical variables. Demographic variables included age and postal code. ED administrative variables included anonymized case identification number, anonymized admission number, and ED registration date. Clinical variables included the presence of 17 comorbidities from the past 5 years of hospital discharge records before the index emergency admission, and primary ED diagnosis. Patients’ identifying information was removed to ensure anonymity.
Identification of pre-existing comorbidities
The definitions of comorbidities employed in this study were based on the Charlson Comorbidity Index . The 17 comorbidities defined in this study included prior myocardial infarction, congestive heart failure, peripheral vascular disease, cerebrovascular disease, dementia, chronic pulmonary disease, rheumatological disease, peptic ulcer disease, mild liver disease, diabetes, cerebrovascular (hemiplegia) event, moderate to severe renal disease, diabetes with chronic complications, cancer without metastases, moderate to severe liver disease, metastatic solid tumour, and acquired immune-deficiency syndrome (AIDS). Pre-existing comorbidities were determined from their past 5 years of hospital discharge records before the referenced emergency admission. The number of pre-existing comorbidities was further grouped into three categories – no pre-existing comorbidity, single pre-existing comorbidity, and pre-existing multimorbidity in which the patient had two or more of the comorbidities. For our dataset, the information needed to trace 5 years back in their medical records was only available from 2012 onwards. Therefore, the timeline for this analysis only included 2012-2017.
SNOMED CT to ICD-10 conversion for primary ED diagnosis
Our eHints dataset recorded primary ED diagnoses according to the International Classification of Diseases Version 9 (ICD-9) from 2008 to 2014. From 2015 to 2017, the EHR switched to Systematized Nomenclature of Medicine Clinical Terms (SNOMED CT) for the computerized coding of ED diagnoses. To facilitate consistent comparison of primary ED diagnoses across the study period, SNOMED CT codes were first converted to ICD-10 using SNOMED CT to ICD-10-CM Map released by the National Library of Medicine and ICD-10-CM Official Guidelines for Coding and Reporting [25, 26].
Identification of chronic conditions in primary ED diagnosis
To determine whether the primary ED diagnosis would be categorized as a chronic condition or non-chronic condition, we adapted Chronic Condition Indicator (CCI) for ICD-9  and ICD-10 , respectively, developed by Agency for Healthcare Research and Quality. A chronic condition is defined as a condition expected to last 12 months or longer and results in functional limitations and or the need for ongoing medical intervention . Therefore, each emergency admission episode’s primary ED diagnosis can be designated either chronic or not chronic based on its ICD-9 or ICD-10 code.
Identification of ambulatory care sensitive conditions (ACSC) in primary ED diagnosis
Emergency admissions for ACSC were identified from the ICD-9 or ICD-10 primary ED diagnoses. The lists of diagnosis codes adopted in this study were based on the lists described by Billings et al.  and the subsequent ICD-10 version . The algorithm detects 24 ACSC conditions. ACSC conditions are further categorised into acute, chronic, and avoidable conditions. This study focused on the nine chronic ACSC (Table S1).
Data wrangling and analysis were performed using R version 4.0.2 (R Foundation, Vienna, Austria). Chronic conditions, number of pre-existing comorbidities, and chronic ACSCs were analysed as proportions of the total SGH ED admissions. Proportion estimates in a given year were reported with 95% confidence interval. Mann-Kendall test (MK) was used to statistically assess whether there is a monotonic upward or downward trend of a proportion over time. A monotonic upward or downward trend means the proportion consistently increased or decreased over time, but the trend may not necessarily be linear. Additionally, a modified Mann-Kendall test (MMKH) using Hamed and Rao variance correction approach  was performed on trend analysis to address the issue of serial correlation. The serial correlation was evaluated using the acf function in R. The serial correlation (at lag 0-20) was detected in monthly aggregated data. Thus we performed the MMKH test for the monthly aggregated analysis. For annually aggregated data with much fewer time points, the serial correlation was negligible. Trend analysis was performed on the proportions of elderly emergency admissions with pre-existing multimorbidity at the time of admission, whose primary ED diagnosis was categorized as chronic conditions, and whose primary ED diagnosis was identified as ACSC.