Data and sources
The data related to patients admitted to any of the four major public hospitals in Tasmania, Australia (i.e. Royal Hobart Hospital, Launceston General Hospital, Mersey Community Hospital, and North West Regional Hospital). These four hospitals are collectively responsible for 95% of all public hospital admissions in the Tasmanian state that has 500,000 inhabitants [
26,
27]. In this retrospective study, we extracted de-identified demographic and clinical information for hospital admissions that occurred between 1 July 2010 and 30 June 2015. We accessed the “Admitted patient care National Minimum Data Set (NMDS)” for these admissions as provided by the Tasmanian Department of Health and Human Services. The admitted patient care NMDS is a minimum set of data elements for inpatient care that health authorities around Australia mandatorily report to support national collation [
28,
29]. We used specific fields from this dataset for each admission: hospital; patient sex, date of birth, date of admission and discharge (or separation), area of residence; and a sequence of international classification of disease (ICD)-10 codes extracted from medical records by clinical coders. One of these ICD-10 codes from each admission was designated as the primary diagnosis.
Study measures
We analysed data for admissions where the primary diagnosis was one of five chronic medical conditions: cancer (lung and colorectal only), COPD, diabetes (type 2), IHD or stroke. The next step involved further screening of the patients, to identify those with a mental illness ICD-10 code entered as a secondary diagnosis during the course of the admission. Table
1 shows the ICD-10 codes used to define each of the chronic medical conditions and mental illnesses.
Table 1
Coverage of chronic medical conditions and mental illness in this study
Chronic medical conditions as primary diagnosis | Cancer: Lung cancer “C33”,“C34”; Colorectal Cancer “C18”,“C19”,“C20”,“C21” |
COPD: “J44.” |
Type II Diabetes: “E11” |
IHD: “I20”,“I21”,“I22”,“I23”, “I24”, “I25” |
Stroke: “I60”,“I61”,“I62”,“I63”,“I64”,“I65”,“I66”,“I67”,“I68”,“I69”,“G45”,“G46” |
Mental illnesses as secondary diagnosis | “F04”, “F06”, “F07”, “F09”, “F10–F19”, “F20–F29”, “F30–F39”, “F40–F48”, “F50–F59”, “F60–F69”, “F70–F79”, “F80–F89”, “F90–F99” |
Consistent with previous reporting of mental illness, we omitted patients with dementia (F00-F03) [
31] and delirium (F05) [
31,
32] from the mental illness cohort. Complete information on all variables was available for over 99% of admissions, making it appropriate to exclude a case (i.e. patient admission) from analysis only if the required data for the specific analysis was missing.
Other variables computed from the extracted data were LOS, bed days’ use, age, socioeconomic status (SES), and presence of an additional comorbidity. The measurement of the primary outcome measure, LOS, was in whole days, equal to date of discharge or separation minus admission date. We coded LOS as zero (0) for admissions where discharge date was the same as admission date. Calculation of bed day’s use, which is another outcome variable, was based on the average percentage increase (p) in the adjusted LOS of the negative binomial regression run in this study. Explanation of this calculation is provided in the statistical analysis section of the paper. Average measures of SES for applicable geographical regions were downloaded from the website of the Australian Bureau of Statistics [
33]. This average measure of SES is drawn from the Index of Relative Socio-Economic Advantage and Disadvantage (IRSAD), a weighted average score of an area’s characteristics such as material and social resources [
34]. Data for this index is recorded in the Australian census, standardised to a distribution where the mean equals 1000 and standard deviation is 100 [
34]. The IRSAD score reflects an area’s average socio-economic status rather than that of individuals living in that area [
35]. Higher scores in the index infer higher level of advantage, with areas of higher-than-average socio-economic advantage having scores above 1000.
Charlson comorbidity is a method to identify comorbidities and variation of complexity in patients with a list of 17 comorbidities [
36‐
38]. We used lists of ICD-10 codes validated by Quan et al. [
38] to assess the presence or absence of additional Charlson comorbidities in patients. Patients were dichotomised into group of higher complexity when additional Charlson comorbidities were present and vice versa for the group of low complexity. For four of the five chronic medical conditions listed in Table
1 (Type II diabetes, cancer, stroke and COPD), all patients with the condition had at least one of the 17 Charlson comorbidities. Accordingly, we classified patients with these four conditions for presence or absence of additional Charlson comorbidities other than the Charlson comorbidity associated with the primary condition. For example, patients with a primary diagnosis of Type II diabetes were classified into those with and without a Charlson comorbidity other than “diabetes without chronic complication”. Similarly, patients with a primary diagnosis of (a) cancer, (b) stroke and (c) COPD were classified into groups of those with and without a Charlson comorbidity other than (a) “malignancies including lymphoma and leukaemia except malignant neoplasm of the skin”, (b) “cerebrovascular disease” and (c) “chronic pulmonary disease”, respectively. For the fifth condition in Table
1 (IHD), even though there are several cardiovascular comorbidities among the 17 Charlson comorbidities by definition, not all patients with a primary diagnosis of IHD had one of these comorbidities. Hence, patients with a primary diagnosis of IHD were divided simply into those with at least one Charlson comorbidity and those without any Charlson comorbidities.
Statistical analysis
To address the first research question, we utilised descriptive statistics to compare patients within each chronic medical condition, for scenarios of with and without the comorbidity of mental illness. T-tests were used to compare the means of continuous variables, and Fisher exact tests were used to compare distributions of binary variables.
We addressed the second research question with stages of analyses. Firstly, negative binomial regression was used to test for an independent association between LOS and comorbidity of mental illness. This analysis was run using the glm.nb() command from the “MASS” package in R, which contains functions and datasets associated with Venables and Ripley's Modern Applied Statistics With S [
39]. The analysis was adjusted for six potential confounding factors: hospital, financial year of separation or discharge (categorical), presence of Charlson comorbidity, age (5-year categories), gender and SES (treated as a continuous variable). To test whether different subtypes of mental illness were associated with different changes in LOS, the mental illness diagnoses in Table
1 were divided into their blocks as per ICD-10 WHO version [
40]: F04–F09 (excluding F05), F10–F19, F20–29, F30–39, F40–F48, F50–F59, F60–F69, F70–F79, F80–F89, F90–98 and F90–F99. In the model for each chronic condition, if codes from a particular block occurred in 10 or more admissions, that block was coded as an additional binary variable (in addition to the binary variable coding of overall presence/absence of mental illness) and added to the regression model. Backwards stepwise regression [
35] was then used to remove non-significant blocks until all blocks remaining in the model (if any) had Wald p-values less than 0.05. Changes in the adjusted LOS were calculated for each remaining significant block. The same calculation was applied for all other sub-type of mental illnesses, that is, the group of non-significant blocks and blocks with less than 10 admissions.
The next stage was assessing the independent association of comorbidity of mental illness with LOS, for differences in usage of bed days. This aspect of the analysis required several steps. First, we estimated the average percentage increase (
p) in adjusted LOS, associated with mental illness (Column 5 in Table
4). This percentage increase
p is equal to (exp(
b) – 1) × 100%, where
b is the coefficient of the mental illness term in the negative binomial regression model. Then, we calculated the difference in bed days’ use, reducing the p to zero, with the assumption of same average LOS for patients of chronic medical conditions, between the scenarios of with and without a comorbidity of mental illness. For example, if we consider n patients having a comorbidity of mental illness with a particular chronic disease (Column 1 in Table
4) and that average LOS for these n patients is t days (Scenario 2 in Column 2 in Table
4). Then, the estimated difference in bed days’ use in five years came to be n*t*p / (1+ p) (Column 6 in Table
4). Similarly, the estimated bed days’ use per patient came to be t*p / (1 + p) (Column 6 in Table
4). The estimated bed days’ use represents the adjusted value, controlling for the influence of the confounding variables.