Increased 30-day mortality associated with weekend emergency admission for alcohol-related liver disease in England: a record-linkage study using the Clinical Practice Research Datalink

Abstract

Objective

Deficiencies have been highlighted in acute hospital care for alcohol-related liver disease (ARLD). Such problems may be worse at weekends (WEs). Increased 30-day mortality for WE admissions has been reported for several acute conditions, but data for ARLD are limited. We aimed to compare patient and pathway characteristics between WE and weekday (WD) admissions and investigate the ‘weekend effect’ on mortality.

Methods

Retrospective cohort study (2008–2018) using linked electronic databases (Hospital Episode Statistics-Clinical Practice Research Datalink and death registration) including 17 575 first emergency admissions identified using the Liverpool ARLD algorithm. Exposure: WE admission (Saturday or Sunday). Main outcome: all-cause death within 30 days. Covariates included socio-demographic characteristics, pathway characteristics (pre-admission contacts and admission method) and markers of severity (recorded stage of liver disease, ascites and varices, comorbidity). Alternative risk-adjustment methods were used, including standard regression and propensity-weighted analysis (Inverse Probability of Treatment Weighting).

Results

3249 admissions (18.5%) were at WE. Unadjusted 30-day mortality was significantly higher for WE versus WD (17.1% vs 15.5%, p=0.018). All models demonstrated increased odds of death for WE admissions with adjusted ORs ranging from 1.15 to 1.23 (relative risk of 1.12–1.19). Causes of death did not vary by admission day and effect was consistent across subgroups. Findings were robust to sensitivity analyses restricting the cohort to patients admitted directly from Accident and Emergency department (A&E), or cirrhosis or ascites but not varices.

Conclusion

First ARLD admissions at the WE experienced a 12–19% increase in 30-day mortality risk compared with WD. Although residual confounding cannot be excluded, this suggests the possibility of avoidable mortality among those hospitalised at WEs. Services should be alert to risks of WE effects when planning care.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• A ‘weekend effect’ on mortality has been reported for emergency admissions overall and in several specific conditions but the extent to which this phenomenon applies to alcohol-related liver disease is unclear.

WHAT THIS STUDY ADDS

• In England between 2008 and 2018, people hospitalised as an emergency for their first time with alcohol-related liver disease at the weekend were at increased risk of short-term mortality. The relative risk of 30-day death was 12–19% higher than those admitted on weekdays. These findings were robust to a range of cohort restrictions and multiple risk-adjustment approaches.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• Healthcare teams and commissioners should be alert to the possibility of weekend effects in this patient group when organising and delivering care, auditing local practice and benchmarking services.

Introduction

Multiple studies have reported an adverse effect of weekend (WE) hospital admission on mortality, including all-cause emergencies and specific acute medical conditions.¹,³ The reason for such associations could be explained by differences in illness severity (ie, sicker patients being admitted at WEs) and/or shortcomings in care provided over WEs when services may be reduced.⁴ The latter interpretation has fuelled intense debate in the UK National Health Service about providing 7-day access to acute specialist care.⁵

Alcohol-related liver disease (ARLD) is an increasing global problem with rising prevalence.⁶ Emergency admission for ARLD is associated with high mortality. Numerous reports have highlighted concerns about the quality of care delivered to acutely hospitalised patients during the crucial first hours or days of admission.^{7 8} Early recognition and treatment of specific complications is vital.^{9 10} We have reported recently that one in seven patients died during their first unplanned admission for ARLD in England between 2008 and 2018, with only small improvements in prognosis over that decade.¹¹

In the present study, we tested the hypothesis that patients admitted over WEs for ARLD experience an increase in 30-day mortality. There are very few reports in the literature addressing this question. We used linked data from primary and secondary care to take account of pathway characteristics and an algorithm-based method for characterising case-mix with condition-specific markers of severity.¹¹,¹³

Methods

Data sources

The Clinical Practice Research Datalink (CPRD) provided access to linked data from primary and secondary care (Hospital Episode Statistics, HES) and to death certification (Office for National Statistics). CPRD is an anonymised database covering a broadly representative sample of the English population (over 11 million patients from more than 650 practices).¹⁴ The primary care database contains medical terms recorded by the general practitioner using Read codes. HES contains administrative information and diagnostic codes generated by clinical coders after discharge from NHS hospitals using the International Classification of Diseases (ICD-10) system. HES linkage was available for the period April 1997 to March 2021.

Case definition and cohort selection

The method for extracting this patient cohort from CPRD has been reported previously.¹¹ Briefly, cases were registered with a CPRD practice in England (≥1 year) and had experienced an index emergency admission for ARLD over a ten year period between 1 April 2008 and 31 March 2018. Eligible admissions were identified and classified using the Liverpool alcohol-related liver disease algorithm (LAA).¹¹,¹³ The LAA case definition includes a ‘primary’ subgroup having a primary discharge diagnosis of ARLD based on specific codes (K70.0, 70.1, 70.2, 70.3, 70.4 or 70.9) and an ‘uplift’ subgroup where there is a combination of primary and secondary codes compatible with emergency admission for ARLD.¹¹,¹³ The uplift coding patterns allow for ARLD-specific codes to be recorded as a secondary diagnosis, provided there is also a primary diagnosis consistent with a symptom, sign or specific complication of ARLD or another associated alcohol-specific condition.^{12 13} The development and validation of the LAA has been reported previously.¹¹,¹³ We defined index emergency admission as having had no unplanned admission for ARLD within the preceding 10 years, as previously described.¹¹

Patient characteristics

Patient-level covariates included socio-demographic data (age group, sex, deprivation status, region of residence (10 regions of England)). A binary variable was created to indicate primary or uplift discharge coding pattern based on the LAA.¹¹,¹³ The recorded stage of liver disease was extracted, based on relevant ICD-10 codes (including unspecified stage, K70.9), as described.¹¹ Additional markers of advanced disease were based on the full list of codes recorded on discharge, creating binary flags for ascites and varices.¹¹ A categorical variable for non-liver comorbidities was created, based on the Charlson Comorbidity Index but with omission of the liver-related codes.¹¹

Although specific codes for hepatic encephalopathy are included in the LAA for case identification, such codes were recorded rarely in this cohort (2.5% overall), and were found very rarely in the absence of codes for varices or ascites (<1%). As previously reported, it is possible that encephalopathy may be coded using non-specific ICD-10 codes for confusional states or coma, but these lack the specificity of ascites or varices for liver-related complications and could be caused by alcohol intoxication, withdrawal or other unrelated causes.¹⁰ Hence, we did not assign a separate case-mix variable to flag the small minority of patients with specific codes in the main analyses.

Pathway characteristics

We examined records of general practice (GP) consultations within the year prior to admission and screened for the presence of any Read codes relating to alcohol or liver disease. The relevant code lists have been published previously.¹¹ We categorised patients according to whether or not they had ≥1 GP contact in the year before index admission, and (if so) whether any contacts were alcohol-related or liver-related. This assigned cases into one of three groups: (1) no prior consultation record; (2) ≥1 consultation with alcohol or liver problems recorded; (3) ≥1 consultation but without any alcohol or liver problems recorded.

Similarly, we screened HES data for all emergency admissions within the prior year and created variables to indicate whether or not the patient had experienced a prior admission and (if so) whether an alcohol-specific code was recorded on discharge. Again, this resulted in allocating cases into three groups: (1) no prior emergency admission; (2) ≥1 admission with an alcoholic-specific code recorded; (3) ≥1 admission but without an alcoholic-specific code recorded.

The method of hospital admission is recorded in HES and includes a number of categories. The vast majority of emergency admissions are either directly via the (A&E) or following an urgent referral for hospital admission from GP services. We categorised admission method using a binary variable to distinguish admissions that did not involve prior contact with another service (ie, directly via A&E) from those that did (any other route).

Exposure: weekend admission

WE admission was defined as having a recorded date of admission falling on a Saturday or Sunday.

Outcome: 30-day all-cause mortality

The primary outcome for the present study was deaths within 30-days of index admission (in or out-of-hospital) based on linkage to death certification.

Statistical methods

First, we compared patient characteristics between WE and weekday (WD) admissions. Categorical variables are reported as counts and percentages (compared using χ² statistics). Factors associated with WE admission were further identified using binary logistic regression (BLR).

Second, we compared characteristics of 30-day deaths with survivors and identified factors associated with death. Crude case fatality rates (CFRs) were calculated as the count of deaths with 30 days (numerator) divided by number of first admissions (denominator). Third, we tested for differences in crude CFRs between WD and WE admissions for the cohort overall and also stratified by recorded stage of liver disease.

Fourth, we fitted traditional BLR models to test for an independent association between WE admission and 30-day death. A series of alternative models was used, as summarised in table 1. These included an unadjusted model (exposure and outcome only), a simple model (using only traditional generic covariates) and then more complex disease-specific models which included various markers of illness severity, prior healthcare contacts and method of admission. Covariate selection was based on clinically-plausible associations, including case-mix variables shown to be associated with increased risk of in-hospital mortality in this same cohort.¹¹ All models were adjusted for year of admission and region of residence, as previously described.¹¹ We categorised age into age-bands for ease of interpretation of ORs, but our study findings were unchanged in sensitivity analyses using age as a continuous variable in each of the models. We created an additional flag for public holidays in England (Bank Holiday Mondays), allowing their exclusion in sensitivity analyses.

Table 1. Summary of models and risk-adjustment covariates (predictors) used to test the association between weekend admission and 30-day mortality among index admissions for alcohol-related liver disease.

Model	Method	Covariates (predictors)
1 – unadjusted	BLR	Exposure only (weekend admission)
2 – generic covariates	BLR	Age group Sex Charlson Comorbidity Index Deprivation quintile Region of residence Year of admission
3 – liver-specific covariates	BLR	Age group Sex Non-liver comorbidities Deprivation quintile Primary or uplift coding pattern Recorded stage of liver disease Coding for ascites Coding for varices Region of residence Year of admission
4 – with pathway characteristics	BLR	Model 3 plus: Prior GP consultations Prior emergency admissions Admission method
5 – IPTW	Weighted BLR	Propensity weights based on: Age group Sex Prior GP consultations Prior emergency admissions Admission method Primary or uplift coding pattern Non-liver comorbidities Recorded-stage of liver disease Coding for ascites Coding for varices See figure 1 Outcome model: Deprivation quintile Region of residence Year of admission

Models 1 to 4 are binary logistic regression models with the stepwise addition of generic, liver-specific and pathway-related covariates. Model 5 uses a propensity weighting method (IPTW) to account for covariates associated with both admission timing and outcome, then adjusts for any additional factors associated with outcome alone.

BLR, binary logistic regression; GP, General Practice; IPTW, Inverse Probability of Treatment Weighting.

Fifth, for Inverse Probability of Treatment Weighting (IPTW) analyses, we used a BLR model to generate inverse probability weights.¹⁵ The propensity model assigned WE admission the dependent variable and the predictors were those covariates expected to be associated with both exposure (WE admission) and outcome (30-day death).¹⁶ Covariate selection was based on theoretically plausible causal pathways, summarised in figure 1. The outcome was then modelled using a weighted-BLR which included the inverse-probability weights derived from the propensity model plus any additional covariates associated with outcome alone, year of admission and region of residence (table 1). In sensitivity analysis, we restricted the propensity model to include only those factors demonstrating significant univariate associations with exposure and outcome in the current cohort (p values of ≤0.2). For ease of interpretation of the ‘effect’ size, we also provide an estimate of relative risk (RR) derived from ORs.¹⁷

Figure 1. Theoretical causal associations between baseline patient characteristics, weekend admission (exposure) and 30-day death (outcome) following index admission for alcohol-related liver disease. Directed acyclic graph. The aim of the study was to estimate a possible causal effect of weekend admission (green node) on outcome (blue node) after accounting for confounders (red nodes). Many of the available covariates can be regarded as surrogate markers of overall severity of illness at the time of admission (grey node), which in turn is a key factor in determining outcome. Some of the same factors may influence likelihood of weekend admission directly (lines not drawn) or via an association with illness severity. There remains the possibility of residual confounding from factors not considered. AED, admitted directly via accident and emergency department. GP, General Practice.

Associations between WE admission and 30-day death are presented in the order of models described in table 1, starting with unadjusted (crude) analyses (Model 1); then adjusting for the generic covariates (as typically reported in the literature on WE effects; Model 2); then adding condition-specific surrogate markers of disease severity (Model 3); and finally pathway characteristics (Model 4). Lastly, we present findings from the IPTW approach as an alternative to testing the study question (Model 5).

All analyses were undertaken in R (V.4.1.2, R Core Team 2021, R Foundation for Statistical Computing, Vienna, Austria), on a complete case basis with minimal missing data (n=19 (0.1%) for region of residence in the original cohort extract).¹⁰ Results are reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology checklist (see: online supplemental appendix 1).¹⁸

Results

Overall characteristics, yearly trends and geographical distribution for this cohort have been described previously.¹¹ Of the 17 575 index admissions, 3249 (18.5%) were at WEs (table 2). The volume of daily admissions (incidence) was significantly lower on both a Saturday and a Sunday compared with WDs (figure 1).

Table 2. Socio-demographic, pathway and clinical characteristics of 17 575 first emergency admissions for alcohol-related liver disease (English CPRD population, 2008–2018), stratified by day of admission.

Characteristic	Overall (n=17 575)	Weekday (n=14 326)	Weekend (n=3249)	P value
Age group				0.10
<40	2326 (13.2%)	1856 (13.0%)	470 (14.5%)
40–49	4534 (25.8%)	3675 (25.7%)	859 (26.4%)
50–59	5137 (29.2%)	4208 (29.4%)	929 (28.6%)
60–69	3839 (21.8%)	3160 (22.1%)	679 (20.9%)
70+	1739 (9.9%)	1427 (10.0%)	312 (9.6%)
Sex				0.084
Female	5820 (33.1%)	4786 (33.4%)	1034 (31.8%)
Male	11 755 (66.9%)	9540 (66.6%)	2215 (68.2%)
Deprivation quintile				0.6
1 (least deprived)	2435 (13.9%)	2007 (14.0%)	428 (13.2%)
2	2735 (15.6%)	2215 (15.5%)	520 (16.0%)
3	3028 (17.2%)	2449 (17.1%)	579 (17.8%)
4	3837 (21.8%)	3140 (21.9%)	697 (21.5%)
5 (most deprived)	5540 (31.5%)	4515 (31.5%)	1025 (31.5%)
Prior GP consultation in last year (≥1)				0.010
None	7639 (43.5%)	6216 (43.4%)	1423 (43.8%)
Alcohol or liver codes recorded	5396 (30.7%)	4462 (31.1%)	934 (28.7%)
Alcohol or liver codes not recorded	4540 (25.8%)	3648 (25.5%)	892 (27.5%)
Prior emergency admission in last year (≥1)				0.001
None	10 310 (58.7%)	8483 (59.2%)	1827 (56.2%)
Alcohol	5525 (31.4%)	4417 (30.8%)	1108 (34.1%)
Other	1740 (9.9%)	1426 (10.0%)	314 (9.7%)
Admission method				<0.001
Directly via A&E	12 721 (72.4%)	9885 (69.0%)	2836 (87.3%)
Any other route	4854 (27.6%)	4441 (31.0%)	413 (12.7%)
Case definition				<0.001
Primary	9374 (53.3%)	7804 (54.5%)	1570 (48.3%)
Uplift	8201(46.7%)	6522 (45.5%)	1679 (51.7%)
Charlson Comorbidity Index score				0.3
0	6499 (37.0%)	5290 (36.9%)	1209 (37.2%)
1–10	4960 (28.2%)	4079 (28.5%)	881 (27.1%)
>10	6116 (34.8%)	4957 (34.6%)	1159 (35.7%)
Non-liver comorbidity score				0.7
0–1	15 506 (88.2%)	12 632 (88.2%)	2874 (88.5%)
2+	2069 (11.8%)	1694 (11.8%)	375 (11.5%)
Recorded-stage of liver disease				0.3
Fatty liver	942 (5.4%)	768 (5.4%)	174 (5.4%)
Hepatitis	2369 (13.5%)	1912 (13.3%)	457 (14.1%)
Fibrosis and sclerosis	62 (0.4%)	48 (0.3%)	14 (0.4%)
Cirrhosis	5001 (28.5%)	4104 (28.6%)	897 (27.6%)
Hepatic failure	2354 (13.4%)	1946 (13.6%)	408 (12.6%)
Not specified	6847 (39.0%)	5548 (38.7%)	1299 (40.0%)
Coding for ascites	6589 (37.5%)	5595 (39.1%)	994 (30.6%)	<0.001
Coding for varices	2704 (15.4%)	2173 (15.2%)	531 (16.3%)	0.094

Patient characteristics

Those first-hospitalised at WEs were very slightly younger (mean age: 53.0 vs 53.5 years, p=0.048), with a marginally higher proportion of men (68.2% vs 66.6%, p=0.087) but lower proportions with primary coding pattern (48.3% vs 54.5%, p<0.001) or coding for ascites (30.6% vs 39.1%, p<0.001). However, there were no significant differences with respect to deprivation status, non-liver comorbidity, recorded stage of liver disease or coding for varices (table 2).

Pathway characteristics

Just over half the cohort had a record of ≥1 GP consultation in the year preceding index admission, but this did not differ by day of admission (table 2). Among those with prior contacts in primary care, there was a slightly lower proportion with alcohol or liver-related codes recorded among WE admissions (51% vs 55%, p<0.01). WE admissions for ARLD had a slightly higher overall proportion with ≥1 prior emergency hospitalisation in the past year (44% vs 41%, p<0.01). The overall proportion with a prior admission containing alcohol-specific coding was marginally higher for WE hospitalisation (34% vs 31%, p<0.01). However, the most notable difference in pathway characteristics was in relation to admission method. As expected, the proportion of WE admissions admitted directly via the A&E department (as opposed to any other route, such as GP referral) was significantly greater (87.3% vs 69.0%, p<0.001). The univariate associations between patient or pathway characteristics and WE admission (table 2) are further summarised as ORs in online supplemental table S1.

Characteristics associated with 30-day mortality

Overall, there were 2771 deaths within 30-days of index admission (crude CFR: 15.8%). Among 30-day decedents, the median time-to-death was shorter for those originally admitted at the WE compared with WDs (7 days vs 9 days, p<0.05 Mann-Whitney U test). The corresponding distributions of time-to-death are shown in online supplemental figure S1, with a ‘peak’ of deaths occurring within 2 days of admission for both groups.

Patient characteristics independently associated with 30-day death were increasing age, female sex, primary case definition, non-liver comorbidity score, recorded stage of liver disease (more advanced stage), coding for ascites and coding for varices (online supplemental tables S2 and S3).

Association between weekend admission and 30-day mortality

Unadjusted and stratified analyses

The crude CFR was significantly higher for WE admissions (17.1% vs 15.5%, p=0.017). Crude rates were higher at WEs for each of the main recorded stages of liver disease (figure 2), but this reached statistical significance only for cirrhosis (19.3% vs 15.7%, p=0.009). The crude rate was significantly higher at WEs for patients with coding for ascites (24.1% vs 20.0%, p=0.003) but not varices (19.6% vs 18.2%, p=0.5). Additional stratified analyses are shown in online supplemental figure S2, comparing crude rates for WD versus WE admissions across age, sex, comorbidity and pathway categories.

Figure 2. Crude case fatality rate (30-day) for weekday versus weekend index emergency admissions by recorded stage of alcohol-related liver disease. Error bars show 95% binomial confidence limits. The crude rates were numerically higher for all main stages, but this was only significant for cirrhosis (p=0.009). CFR, case fatality rates.

Conventional case-mix adjustment models

Each of the alternative modelling approaches identified a significantly increased odds of death for WE admissions (table 3; online supplemental figure S3). Compared with the traditional risk-adjustment method based on generic covariates (Model 2, OR: 1.15 (95% CI: 1.04 to 1.28)), the use of disease-specific severity markers increased the magnitude of the observed association (Model 3, OR: 1.23 (95% CI: 1.10 to 1.37)). Adding pathway characteristics retained an independent association for WE admission (Model 4, OR: 1.15 (95% CI: 1.03to 1.28)). Model diagnostics suggested that Model 4 was the best fit for the data (table 3).

Table 3. Association between weekend admission and 30-day all-cause mortality for first emergency admissions for alcohol-related liver disease.

Model	OR (95% CI)	P value	C-statistic*	AIC†	Relative risk‡
1 – unadjusted	1.13 (1.02 to 1.25)	0.017	0.51	15 316	1.11
2 – generic covariates	1.15 (1.04 to 1.28)	0.009	0.68	14 393	1.12
3 – liver-specific covariates	1.23 (1.10 to 1.37)	<0.001	0.72	13 866	1.19
4 – with pathway characteristics	1.15 (1.03 to 1.28)	0.014	0.73	13 777	1.12
5 – IPTW§	1.18 (1.09 to 1.28)	<0.001	N/A	N/A	1.15

ORs and estimated relative risk derived from binary logistic regression models fitted with alternative sets of covariates as summarised in table 1. Patients admitted at the weekend experienced a 12–19% increase in risk of dying within 30 days compared with those admitted on a weekday.

Adding additional predictors significantly improved model fit from Model 2 to Model 3 (χ2=541; p<0.001), and from Model 3 to Model 4 (χ2=98; p<0.001). This indicates that the addition of liver-specific covariates added substantial improvement in explanatory power compared with Model 2, whereas further addition of pathway characteristics to Model 3 added a smaller but still significant improvement. Hence, Model 4 represented the ‘best fit’ of the traditional BLR models.

^*Values for the C-statistic range from 0 to 1, with higher values indicating better goodness-of-fit for the model.

^†AIC=Akaike’s information criteria. A lower value indicates a better model ‘fit’ relative to models fitted on the same dataset that have higher AIC values.

^‡Relative risk estimated from the adjusted OR.

^§IPTW=Inverse Probability of Treatment Weighting (see Methods). Likelihood ratio tests were conducted to compare the nested BLR models (Models 2, 3 and 4).

BLR, binary logistic regression.

Using Model 4 we confirmed that the independent association between WE admission and 30-day mortality remained significant when the cohort was restricted to only those with a ‘primary’ case definition (n=9374; OR: 1.21 (95% CI: 1.05 to 1.40), p=0.009), those with a recorded stage of alcoholic cirrhosis only (n=5001, OR: 1.27 (95% CI: 1.04 to 1.54), p=0.017), those admitted directly from A&E (n=12 721; OR: 1.13 (95% CI: 1.01 to 1.27), p=0.04) or those with coding for ascites in their discharge record (n=6589; OR: 1.19 (95% CI: 1.01 to 1.41), p=0.039). This was not the case for the subgroup with varices recorded (n=2704; OR: 1.08 (95% CI: 0.84 to 1.39), p=0.5) although this may reflect the smaller sample size with reduced power for studying this subgroup. Sensitivity analysis excluding admissions on a Bank Holiday Monday from the cohort revealed no differences in any of the main study findings (data not shown).

Inverse Probability of Treatment Weighting

As an alternative to conventional case-mix adjustment, we also applied a propensity-score based method (IPTW) to assess the ‘weekend effect’. All factors theorised a priori to be associated with both WE admission and 30-day death were used to generate relevant weights (figure 1 and table 1), adding any remaining covariates to a final outcome model. Our base case IPTW analysis yielded an OR of 1.18 (95% CI: 1.09 to 1.28; p<0.001) as the estimated causal WE effect on mortality. Supporting data demonstrating the successful balancing of covariates after IPTW is summarised in online supplemental table S4. A further sensitivity analysis was conducted whereby the propensity model contained only those factors which demonstrated significant associations with WE admission in the dataset (age, prior GP admission, admission method, case definition, stage of liver disease, ascites and varices), adding the remaining covariates to the weighted outcome model. This produced the same results.

Weekend effect across patient subgroups

In additional analyses, we explored whether the observed ‘weekend effect’ differed across patient subgroups. To do so, we included interaction terms between WE admission and either age group, sex, comorbidity category or pathway characteristics in the original logistic regression model (Model 4). Likelihood ratio tests comparing models with, and without, each interaction showed no significant interaction for age group (p=0.98); sex (p=0.11); non-liver comorbidity (p=0.08); those with/without prior emergency admission (p=0.126) or a GP contact in the last year (p=0.51). These analyses indicate that the effect of WE admission on 30-day mortality was not significantly modified by these factors. Additionally, we conducted exploratory stratified adjusted analyses (online supplemental table S5). The ORs for WE admission within some strata fell below statistical significance, but these post hoc analyses are limited by reduced sample sizes and loss of power. However, the direction and magnitude of the WE effect were largely consistent across all strata, supporting the results of the interaction analysis.

Characteristics of 30-day deaths stratified by weekday versus weekend admission

To explore possible reasons for differences in mortality between WD and WE admissions, we first compared the demographic and clinical characteristics of patients who died within 30 days of WD admission versus those who died within 30 days of WE admission (online supplemental table S1). This stratified analysis showed no differences with respect to age, sex, deprivation status, comorbidity, recorded stage of liver disease or coding for varices. The only significant difference was the proportion of 30-day deaths with ascites coded which was lower (not higher) for WE admissions compared with WDs (43.1% vs 50.5%, p=0.002).

Second, we analysed death certification to verify whether causes of death were comparable between WD and WE admissions. The proportion of death certificates that recorded both liver disease and an alcohol-related cause for 30-day death within any field was 95% for WD (2095 of 2214 deaths) and 95% for WE admission (529 of 557 deaths). In order of frequency, the most common individual ARLD-specific codes recorded as the primary cause of death were: alcoholic liver disease, unspecified (n=647; 23% of total 30-day deaths), alcoholic hepatic failure (n=559; 20% of deaths), alcoholic cirrhosis of liver (n=344; 12%) and alcoholic hepatitis (n=148; 5.3%) and there were no statistically significant differences in proportions of any of these individual certified causes between WD and WE admissions.

We further screened all available fields within the death certificate dataset (ie, the primary cause and up to ten additional cause fields) for additional terms consistent with septicaemia, renal failure or any form of gastrointestinal bleeding—reflecting three common life-threatening complications of ARLD which might be contributors to excess mortality. No significant differences were observed between WD and WE admissions in the proportions of death certificates with these causes listed, with rates of 9.3% (n=205) versus 11% (n=63) for septicaemia (p=0.14): 15% (n=341) versus 17% (n=93) for renal causes (p=0.5); and 15% (n=324) versus 13% (n=72) for bleeding (p=0.3). Only a small proportion of death certificates had varices recorded explicitly as a contributory cause of death but rates were comparable for WD and WE admissions (4.0% vs 3.9% of certificates, p>0.9). Ascites was only rarely recorded on death certificates per se, but again with no difference (0.4% vs 0.2%, p>0.7). Hence, whereas these death certificates almost invariably recorded terms consistent with ARLD, the additional recording of common complications was quite uncommon, but we found no difference between WE and WD admissions to suggest systematic difference in causes of death.

Discussion

Our study focused on whether there was evidence for excess WE mortality in people admitted as an emergency with ARLD for their first time. We found that crude case fatality was significantly higher for WE admissions overall and that the odds of death were consistently increased using a range of alternative risk-adjustment approaches. Expressed as RR, and based on the lowest limit of our estimates, patients admitted at the WE were at least 12% more likely to die within 30 days of first admission. Although age, sex, comorbidity and pathway characteristics were each associated with 30-day mortality in our cohort, we found no evidence that these variables modified the increased risk associated with WE admission. The lack of interaction suggests that this ‘weekend effect’ was consistent across a broad range of patient subgroups, reinforcing the robustness of the observed association.

Our finding of reduced overall volume (incidence rate) of emergency admissions for ARLD on WE days is consistent with studies of other acute conditions.¹,⁴ The reasons for this phenomenon are likely to be multiple, including differences in consultation behaviours, access to care or thresholds for admission. However, the key question is whether WE admissions tend to be selectively ‘sicker’. Interestingly, we found that case mix profiles of WE and WD admissions were very similar, with little evidence of any meaningful difference in generic or condition-specific markers of illness severity. Most of these case mix factors are expected to operate via severity of illness in the causal chain of mortality (figure 1). Notably, we observed similar reductions in volume of admissions across the spectrum of recorded stages of liver disease (figure 3). A selective reduction in milder (earlier) stages would be expected if reduced WE admissions reflect a ‘triage’ effect. Whatever diverse factors were driving lower numbers of admissions at WEs in our cohort, this phenomenon seemed to be operating across the mortality risk continuum.

Figure 3. Distribution of 17 575 index emergency admissions for alcohol-related liver disease by day of admission and stratified by recorded stage of disease. The volume of admissions was lower at weekends, but there was no significant difference in the distribution of recorded stage of liver disease.

We found that certain pathway characteristics were associated with short-term mortality risk for index hospitalisations and adjusted for these effects. GP consulters in the year before admission who had coding for alcohol-related or liver-related problems were at slightly lower risk of 30-day death than those without any consultation record, whereas mortality risk was higher in those consulting for other reasons. Similarly, those patients with prior emergency admission that contained alcohol-specific diagnoses were at slightly lower risk of 30-day death following their subsequent first admission for liver disease. There are various potential reasons why prior alcohol-related health service contacts might be associated with slightly better outcomes for a first emergency admission. Certain care-seeking behaviours might be associated with being hospitalised at an earlier or less severe point in the patient journey. Prior recognition of alcohol problems by the healthcare system might have triggered interventions that attenuated the risk of a fatal first admission. Whatever the mechanisms, these findings allowed us to adjust for prognostically important differences in pre-admission contacts when examining WE effects.

There are only limited previous data published on the WE effect and mortality for ARLD admissions. Roberts et al included ARLD in their study of 19 categories of gastrointestinal emergencies using hospital administrative data for England and Wales (2004–2012), identifying each patient’s first admission during the observation period based on primary discharge diagnosis (K70.x).¹⁹ They observed a crude mortality rate of 16.9% for ARLD overall, reporting a 26.2% increase in 30-day mortality for WEs—double our figure. This higher estimate may relate to reliance on a small set of traditional generic risk-adjusters (age, sex, socioeconomic status and comorbidities) in contrast to our use of an algorithm-based cohort discovery method, several condition-specific markers of severity and further adjustment for pathway characteristics.

There are some international reports of delayed care processes for patients with specific complications of liver disease admitted at the WE. Myers et al found a small but significant delay in endoscopic intervention among patients admitted with oesophageal varices at the WE, based on the National Inpatient Sample (NIS) in the USA (1998–2005).²⁰ However, their study did not detect any difference in risk-adjusted in-hospital mortality. Similarly, negative findings were reported subsequently for variceal bleeding (NIS, 2009).²¹ Consistent with this previous North American experience, we did not detect a WE effect within our English cohort of index admissions after restricting the analysis to cases with varices recorded, although our negative finding might reflect small sample size rather than a true absence of WE effect. More recently, Gupta et al examined the WE effect in patients admitted non-electively for complications of liver disease (NIS, 2005–2014), finding delays in paracentesis for ascites but no statistically significant association could be demonstrated for in-hospital mortality.²² By contrast, restricting our analysis to patients with coding for ascites did suggest a WE effect on mortality among first emergency admissions.

Key strengths of our study are the focus on first emergency admissions (rather than pooling together readmissions at any stage in the patient journey), the application of an algorithm-based method for cohort discovery and characterisation (as standard methods can miss up to half true admissions),¹¹,¹³ the use of linked data from primary and secondary care to account for healthcare contacts in the year before admission,¹¹ the inclusion of disease-specific prognostic markers and inclusion of propensity score-based methodology (IPTW) as an alternative method to traditional regression. By controlling for variables related to exposure, propensity score matching techniques are said to afford researchers the ability to render a more precise estimate of ‘effects’.²³ In the case of ARLD, the estimated effect sizes were comparable to those derived from simple regression. Furthermore, study findings were not altered in extensive sensitivity analyses where we restricted the cohort on the basis of prognostically important patient or pathway characteristics, nor were differential effects seen across case mix strata.

Our study has important limitations. First, our findings apply to the English CPRD population and may not be generalisable to the whole population, nor to every hospital. Second, using CPRD precludes examination of institutional variation as the HES linkage does not include details of admitting hospital. Hence, we cannot infer the WE effect applies to all providers. We used CPRD in order to derive variables related to prehospital contacts in primary care. Future analysis of full national HES data using our methodology would be of interest and could inform benchmarking metrics. National HES lacks any linkage to GP records, but it does include the data needed for most of the prognostically important covariates identified in our study.

Third, we categorised liver disease severity based on ICD-10 codes reflecting ‘recorded-stage’ of disease and specific complications using our previously reported methods.¹¹,¹³ This system does not allow for a clear distinction between ‘decompensated’ and ‘compensated’ liver disease, nor do the datasets contain physical examination findings or laboratory parameters required for clinical severity indices (eg, Child-Pugh score; Model for End-Stage Liver disease) and its variants; European Consortium for Study of the Liver (EASL) Chronic Liver Failure Consortium scores (CLIF-C-AD [Acute Decompensation] and CLIF-C-ACLF [Acute-on-chronic Liver Failure])).²⁴ Walker et al have reported that the inclusion of common biochemical and haematological parameters explained over one-third of the WE effect for Saturdays and Sundays among all-cause admissions.^{25 26} Assuming this same phenomenon applies to our estimate for ARLD, this would still leave a 9% ‘excess risk’ of 30-day death at WE.

Residual confounding cannot be excluded and we recognise that case mix characteristics may vary between WD and WE admissions and our surrogate markers of severity are imperfect. However, most potential ‘unobserved’ confounders will share a common causal pathway for mortality risk via severity of illness. For example, we could not examine delays in care-seeking or admission pathways as a potential factor, since neither the primary care nor hospital datasets provided granular details about symptom onset/duration or admission delays. Delayed hospitalisation for ARLD is likely to translate into more severe disease on admission. Similarly, our datasets lack information on levels or patterns of alcohol consumption. However, current evidence suggests alcohol consumption increases short-term mortality via its impact on liver disease severity rather than being an independent predictor.²⁷ Nevertheless, we acknowledge that further research is needed.

Previous reports have highlighted deficiencies in organisation, quality or safety of care delivered for ARLD in English hospitals.^{7 8} Improving hospital care remains a key priority for the UK Liver Alliance²⁸ and a recent review described mortality from ARLD as an ‘escalating tragedy’.²⁹ Publication of UK Quality Standards for ARLD in 2023 has further emphasised the need for timely detection and management of complications in acutely-hospitalised patients.³⁰ These standards advocate the use of ‘liver bundles’^{9 10} to guide early management and explicitly mandate review by a clinician trained in hepatology within 24 hours of admission.³⁰ Our findings suggest that WE admissions may be more vulnerable to short-term mortality.

Quality improvement initiatives should consider whether care processes and/or access to expertise vary by day of admission, such as compliance with the various components of ‘liver bundles’.^{10 12 30} We did not seek to explore potential care ‘failings’ that might contribute to excess mortality in the present study (eg, shortfalls or delays in specific components of acute generic or condition-specific care). Such factors are multiple, may vary from case to case and are mostly unrecorded in the datasets. Identifying potential causal associations between delays in specific care processes and avoidable mortality in an individual case is challenging and typically needs expert review of the entire clinical record and an individualised judgement.⁷ Our analysis of death certificates did not suggest any systematic difference in reasons for death between WD and WE admissions, as might occur if excess deaths at WEs were consistently related to shortcomings in a particular component of acute care. The recent ALERT-UK audit³¹ used our coding algorithm for case finding and has collected granular local data on care processes across British hospitals—the results may shed further light on common areas of suboptimal care relevant to WE effects.

In conclusion, patients first hospitalised for ARLD in England at the WE appear to have an increased risk of 30-day death. Hospitals should ensure that acute care quality is consistent across WDs and WEs to mitigate the possibility of avoidable mortality—regardless of whether excess WE mortality is driven by illness severity or shortfalls in WE care. Further research is needed to determine whether residual confounding explains part of the effect and to elucidate causes and solutions.

Data availability statement

No data are available.