Ethical approval

This study was approved by the Public Health England Research Ethics and Governance Group (PHE REGG), ref R&D412, on October 26, 2020. Data were anonymised before analysis, and personal identifiers such as name, address, or NHS number were not available to the research team.

This is a case-crossover study estimating the risk of opioid-related deaths associated with admission to the hospital. Case-crossover studies measure acute “triggering” effects of transient exposures [9]. They make within-subject comparisons in the exposure status when an event occurred (in this study, when someone died after using opioids) with the exposure status at other times. Case-crossover studies are one of a family of self-controlled study designs that only include participants who experienced an event. These designs focus on the timing of an event, in contrast to traditional epidemiological studies that focus on who experiences an event. This study is addressing the question “do hospital admission and discharge trigger opioid-related deaths?” We chose this design because it allows inclusion of a large proportion of cases and is statistically powerful, and it controls confounding more effectively than a cohort design that might compare people who use opioids and are admitted to the hospital with those not admitted.

Study participants

We studied opioid-related deaths among people in England aged 18 to 64 between January 1, 2010 and December 31, 2019 based on the date of death (rather than registration). We defined opioid-related deaths as those with an underlying cause of drug poisoning (using the UK Office for National Statistics definition of drug poisoning [10]: the International Classification of Diseases-10th Revision [ICD-10] codes X40-X44, X60-X64, X85, or Y10-14) and where an opioid is also specified (ICD-10 codes T40.0-T40.4 and T40.6) or if opioid dependence (ICD-10 F11) was the underlying cause of death. Coroners investigate drug-related deaths in England, including analysis of toxicology results. This means that the causes of death in this study have been validated to a greater degree than for most deaths. For simplicity, we refer to these deaths as “fatal opioid overdoses” in the title and author summary, although we use “opioid-related deaths” elsewhere to reflect the difficulty of attributing deaths to one specific drug.

Data were drawn from a database that includes mortality data from the UK Office for National Statistics and hospital records from the national Hospital Episode Statistics database, with probabilistic linkage between the 2 sources using the National Health Service (NHS) number, date of birth, sex, and home address [11]. This database does not include deaths if no linkage is found (most likely because the decedent was never admitted to the hospital). Using published mortality data [12], we estimated that 11.8% of deaths were excluded (Fig 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Derivation of study population.

https://doi.org/10.1371/journal.pmed.1003759.g001

Control days

For each case, we sampled 5 days at random from the period 730 to 28 days prior to death, limiting to the same day of the week as death (Fig 2). The reason for limiting to the same day of the week is that both drug-related deaths and hospital admissions vary by weekday, which may cause confounding. Fewer hospital admissions and discharges occur at the weekend [13], while deaths due to drug poisoning peak on Saturday [14]. We chose the period 730 to 28 days prior to death to avoid control days that are too close to death and may have correlated exposures, while also allowing reasonable exchangeability in the probability of hospital admission. We then observed the exposure status on the control days. In sensitivity analysis, we repeated the study with control days sampled from the periods 365 to 28 days before death and 1,095 to 28 days before death.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Illustration of patient timelines and selection of control windows for 5 participants.

Exposure status at time of opioid-related death is compared with 5 days sampled between 730 and 28 days before death.

https://doi.org/10.1371/journal.pmed.1003759.g002

Exposure status

Control days were classified as (A) currently admitted to the hospital (days 1 to 14 after admission); (B) currently admitted to hospital (15+ days after admission); (C) days 1 or 2 after discharge; (D) days 3 to 14 after discharge; and (E) neither hospitalised nor recently discharged (i.e., time spent in the community).

To understand additional differences in the risk of opioid-related death, we further classified admissions as psychiatric or nonpsychiatric; whether postdischarge risk periods followed discharge against medical advice or planned discharge; and whether postdischarge periods followed admissions of 1 day, 2 to 6 days, or 7+ days. We classified admissions as “psychiatric” if the patient was admitted to a specialist mental health provider or if the lead treatment specialty was recorded as “mental health.” We classified discharges as “against medical advice” if the doctor recorded the discharge method as “self-discharged or discharged by a relative or advocate” [15].

Some deaths occurred in hospital. If a patient was admitted due to drug poisoning (see Fig A in S1 Appendix for detail on how we classified these admissions), we moved the date of death to immediately prior to admission and censored the final admission. If a patient died on the day of discharge and hospital data showed that the patient died in the hospital, the death was assigned to exposure status A or B (currently in hospital), while if the hospital data showed that the patient was discharged alive, exposure status C was assigned (discharged in the past 1 to 2 days).

For 28 individuals, we found discordant data in which hospital discharge dates were recorded long after death dates. These discrepant dates are likely due to failed linkage rather than inaccuracies in hospital discharge or death dates, and we excluded these individuals from the study. A small number of individuals had admissions in the 3 days after death, for example, for organ donation, and these cases were retained.

Statistical analysis

We described the characteristics of deaths included in the study and then used conditional logistic regression (stratified by individual) to estimate the association between hospital admission and opioid-related death. The reference category was time neither admitted to hospital nor recently discharged (category E). We repeated the analysis with stratification by sex. We published a protocol before doing analysis [16]. In our protocol, the primary analysis was a self-controlled case series design and have included results using this approach in Fig D and Table B in S1 Appendix. We chose a case-crossover design instead because it allowed analysis of exposures with varying durations. Analysis was conducted using R version 4.0.3.

This study is reported as per the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guideline [17] (see Table D in S1 Appendix).

Description of cases

The study included 13,609 opioid-related deaths. A total of 9,765/13,609 decedents (71.8%) were male, 12,437/13,609 (91.4%) had “white” ethnicity, the median age at death was 42 (interquartile range [IQR] 35 to 49), and decedents predominantly lived in deprived neighbourhoods. Table 1 summarises the characteristics of cases.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Characteristics of individuals included in the study.

https://doi.org/10.1371/journal.pmed.1003759.t001

Most opioid-related deaths (11,629/13,609, 85.5%) did not occur in the hospital or within 14 days of discharge. A total of 656/13,609 deaths (4.8%) occurred in hospital following admission due to drug poisoning. 236/13,609 deaths (1.7%) occurred during a hospital admission where the patient was admitted for a reason other than drug poisoning. The remaining 1,088/13,609 deaths (8.0%) occurred in the 14 days after discharge. Based on published data, we estimated that our study excluded 11.8% of deaths because the decedent was never admitted to the hospital (Fig 1). This suggests that in the population, 7.1% of opioid-related deaths occur in the 14 days after discharge.

Characteristics of hospital admissions in the 2 years prior to death

A total of 37,570 hospital admissions occurred in the 730 days prior to death, with a median of 1 (IQR 0 to 4) per individual. 3,742/37,570 (10.0%) of admissions ended in discharge against medical advice, and 3,418/37,570 (9.1%) were classified as psychiatric admissions. Table 2 summarises the characteristics of hospital admissions in the 2 years prior to death.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Characteristics of hospital admissions in the 2 years prior to opioid-related deaths in England from January 1, 2010 to December 31, 2019.

https://doi.org/10.1371/journal.pmed.1003759.t002

Results of case-crossover analysis

Days 1 to 14 of hospital admissions had a similar risk of opioid-related death as periods in the community (conditional odds ratio of 1.03; 95% CI 0.87 to 1.21; p = 0.95). Days 15+ of hospital admission were associated with lower risk of opioid-related death (conditional odds ratio of 0.41; 95% CI 0.30 to 0.56; p < 0.001).

The risk of opioid-related death increased substantially after discharge, with a conditional odds ratio of 4.39 (95% CI 3.75 to 5.14; p < 0.001) in days 1 to 2 after discharge and 2.09 (95% CI 1.92 to 2.28; p < 0.001) in days 3 to 14. The risk was higher for people discharged after a psychiatric admission and for people who left the hospital against medical advice. Longer admissions were associated with greater risk of opioid-related death after discharge, and we observed this gradient in days 1 to 2 after discharge and days 3 to 14. Results of the case-crossover analysis are shown in Fig 3. Sex-stratified results suggested similar associations for men and women. Sensitivity analysis with control days sampled from 365 to 28 days before death showed slightly smaller associations between hospital discharge and opioid-related death, and sensitivity analysis with control days sampled from 1,095 to 28 days before death showed slightly larger associations. Full results for stratified analyses and sensitivity analyses are provided in Figs D and E and Table C in S1 Appendix.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Risk of opioid-related death according to time proximity to hospital admission (results of case-crossover analysis).

https://doi.org/10.1371/journal.pmed.1003759.g003

Comparison with other studies

We are aware of one other study that has investigated deaths due to drug poisoning in relation to hospital admissions [18,19]. This study used a cohort of people registered for drug treatment in Scotland and reported the rate of drug-related deaths according to time proximity to hospital admissions. The referent in this study was the mortality rate among people who were never admitted to the hospital, and the results may therefore reflect differences between people who were admitted to the hospital and those who were not. The associations are much greater than in our study; for example, the mortality rate in the 28 days after discharge is 15 times that of individuals who were never admitted. Even periods more than 1 year after discharge have 3 times the rate, suggesting that these differences are unlikely to relate to the hospital admission itself.

Research has shown that other life events are also associated with opioid-related deaths. In particular, studies in several countries have found a high risk of drug-related death immediately after the release from prison [20–23] and a protective effect of opioid agonist therapy during this period [24]. Another example is the first 2 weeks after cessation of community-based opioid agonist therapy [25]. These are times when opioid agonist therapy is interrupted, tolerance is changing, and people may get drugs from a different source or use drugs in different ways.

Strengths and limitations

By studying all opioid-related deaths in England, we were able to include people who have never been in drug treatment, a group that is often excluded from studies of this population. The design also meant that we were able to estimate the absolute number of opioid-related deaths that occurred during hospital admissions (236 over 10 years). It also provided power to observe the risk associated with different types and durations of hospital admission, which would be challenging even with an extremely large cohort study.

While the self-controlled methodology eliminates time-invariant confounders, the results may partially be explained by escalating drug use or deteriorating health over time. It is possible, for example, that someone in poor health or at times of more intense drug use would be more likely to be admitted to the hospital and more susceptible to death after using opioids. We tried to limit this type of confounding by selecting control days no more than 2 years before death. The density of hospital admissions does increase before death, but this increase is gradual (Fig C in S1 Appendix).

Our data did not include detailed information about drug use or enrollment in drug treatment services. We therefore could not confirm that participants were using drugs on the control days. People who used illicit opioids during the past decade in England have mostly been using drugs for many years. For example, in a cross-sectional survey of people who injected drugs in England in 2019, the median duration of drug use was 16 years, and only 6% of participants had injected for less than 2 years [26].

Our use of a national hospital dataset meant that we were able to include all hospital admissions in England, but it also meant that we had limited detail about individual patients. One limitation relates to “discharge against medical advice.” This was a binary variable in our analysis, when in reality there are a range of scenarios where discharge is negotiated between patients and hospital staff. A second limitation relates to our classification of hospital admissions ending in opioid-related death. If an admission ended in opioid-related death and the primary cause of admission was a common complication of opioid poisoning, such as cardiac arrest, then we assumed that the indication for admission was related to opioid poisoning. This is a sensitive approach, and, in some cases, the opioid poisoning may have happened while the patient was in the hospital. Therefore, it is possible that we have underestimated the association between hospital admissions (exposure periods A and B) and opioid-related death.

Although our definition of opioid-related death is widely used, its validity is not known. It can be difficult to determine the cause of death when someone dies suddenly and alone. It is possible that some participants in our study who died shortly after a hospital admission died for reasons related to the admission (such as an acute infection) rather than due to opioid poisoning. If the cause of death is unclear but the individual was known to use illicit drugs, a doctor may assume that the death was primarily due to drug use. This type of misclassification may partly explain the association between hospital discharge and opioid-related death.

Interpretation

We identified 3 reasons why discharge from the hospital may be associated with increased risk of opioid-related death. First, opioid tolerance could reduce during an admission. Animal models suggest that opioid tolerance has a half-life of 6 days [27], supporting reductions in tolerance after admissions of 2 to 6 days or 7+ days (the categories used in our analysis). We also saw increased risk after admissions of only 1 day, suggesting that other mechanisms are also important. Second, opioid agonist therapy may be interrupted or reduced either at admission or discharge (see Box 1). This may further contribute to reduced tolerance and increase the likelihood that patients will use riskier opioids such as injected heroin. Third, patients may have an acute illness that makes them more vulnerable to death after using opioids, particularly respiratory problems such as pneumonias and acute exacerbations of chronic obstructive pulmonary disease. These are common reasons for hospital admission in this population [28] and increase the risk associated with central nervous system depressants. Painful conditions may also be important because they are associated with increased use of illicit and prescribed opioids.

We observed a high risk soon after hospital discharge and estimate that 1 in 14 opioid-related deaths in the population occur during these 2 weeks. The safety of discharge may be improved with better linkage to community drug treatment services, and there is a need for research into interventions that can improve continuity of opioid agonist therapy between community and hospital settings. It is thought that half of people who die after using illicit opioids in England have never been in contact with drug treatment services [29]. Initiation of opioid agonist therapy in the hospital is therefore also important. A randomised trial showed that patients that started on buprenorphine in hospital and referred to a community drug service are more likely to continue with treatment and less likely to use illicit opioids, compared with those assigned to opioid detoxification [30]. Hospitals can also provide advice on reducing overdose risk, such as using small test doses and not using drugs alone, and tools to help reduce risk, such as basic life support training and take-home naloxone [31].

We found variation in the risk of opioid-related death in our detailed exposure periods. In particular, the period after discharge against medical advice was associated with opioid-related death. Discharge against medical advice may happen when a patient is experiencing pain or withdrawal and leaves the hospital to use illicit opioids. The finding that the risk of opioid-related death reduces after day 14 of a hospital admission should be treated with caution, because the subset of admissions that are longer than 14 days may be unusual. For example, these patients may have less severe drug dependence or better controlled pain and be less at risk of opioid-related death for these reasons.

In many countries, the average age of people who use illicit opioids is increasing, and the frequency of long-term conditions is also increasing [32]. People who use illicit opioids do not always seek timely healthcare, in part due to fear of stigma, opioid withdrawal in the hospital, and poor pain management [33]. Hospitals that enable patients to disclose illicit drug use without fear of discrimination will be a central element of accessible and high-quality hospital care for this population.

Conclusions

Discharge from the hospital is associated with an acute increase in the risk of opioid-related death, and 1 in 14 opioid-related deaths in England happens in the 2 weeks after the hospital discharge. This supports interventions that prevent early discharge and improve linkage with community drug treatment and harm reduction services.