Given difficulties in using conventional methods to follow-up recipients of THN kits and peers who may experience an opioid overdose, we used routine data to test the feasibility of identifying a high-risk cohort, postulating that outcomes associated with the intervention would be most visible within this cohort. This chapter summarises this feasibility work, which was based on routine data for the population of Wales for the period from January 2015 to November 2021. Summaries of analyses, undertaken within the SAIL Databank, are subject to the SAIL Databank’s dissemination policies.
Routine data
The combined WDSD files comprise information on n = 5,640,113 individuals. No ALF_PE (unique patient identifier in SAIL) was recorded for n = 504 cases; a further n = 1,740,317 individuals had no recorded residency in Wales during the study window; and a further n = 151 individuals had a recorded date of death indicating death before the study window – although date of death is recorded in ADDE, only ADDE data on or after the beginning of the study window were available, and so could not determine whether people were alive or not at this timepoint. Excluding these individuals leaves for further consideration n = 3,899,140 individuals deemed to be alive and resident in Wales at the start of the study window.
Of these, we then excluded another n = 664,050 individuals with week of birth recorded as on or after 1 December 2003; these individuals would be < 18 years old at the end of the study window. A further n = 7694 individuals were aged 17 or under at their final date of residency in Wales, or at their date of death. Excluding these individuals leaves n = 3,227,396 individuals in our study population.
For our derived cohort of n = 3,227,396 individuals, we extracted basic demographic data and then assessed mortality and healthcare resource utilisation, as recorded in the available routine data sources. We start with mortality, which defines both our primary outcome (death from opioid overdose) and study end dates.
Baseline characteristics
Mortality and date of death
Deaths within the study window are recorded in both WDSD and ADDE (see Table 2); we note that WDSD records only the date of death, while ADDE also records cause of death. We regard ADDE data as primary, noting generally good agreement between these datasets.
Combination A: there are n = 233,495 cases where both WDSD and ADDE record death on or before 30 November 2021, with good agreement on the date of death – exact agreement in over 99% of cases. Where there are discrepancies, we take the ADDE date as primary unless that date of death is contradicted by records of subsequent health events, defined as a recorded health event extending beyond the day after the ADDE date of death.
Combination B: there are n = 101 cases where ADDE records a date of death, but WDSD does not. We consider three possible scenarios.
- B1: ADDE date of death is 14 days or more after the last recorded date of residency in Wales.
- B2: ADDE date of death is within 14 days of the last recorded date of residency in Wales.
- B3: ADDE date of death is more than 14 days before the last recorded date of residency in Wales.
For cases in B1, we censor at the last recorded date of residency in Wales; no further data processing is necessary. For cases in B2 and B3, we use the ADDE date of death to define the final date in the study for these individuals, except where that date is contradicted by records of subsequent health events.
Combination C: there are n = 9997 cases where WDSD records a date of death but ADDE does not. We consider two possible scenarios.
- C1: WDSD date of death is 14 days or more after the last recorded date of residency in Wales.
- C2: WDSD date of death is within 14 days of the last recorded date of residency in Wales.
For cases in C1, we censor at the last recorded date of residency in Wales; no further date processing is required. For cases in C2, we use the WDSD date of death to define the final date in the study for these individuals, except where that date is contradicted by records of subsequent health events. Cases in C2 are deaths still to be confirmed within ADDE, and we have no further information available on the cause of death – specifically, we are unable to classify death as opioid-related or otherwise. We note that including these as non-opioid deaths will dilute further the relatively small number of deaths ascribed to opioid overdose. Further dilution will occur on categorising as opioid-related any health events associated with these cases.
Combination D: there is no date of death recorded for n = 2,983,803 cases. For these individuals, we define the study end to be the earlier of the end of the study window or their final recorded date of residency in Wales.
Mortality due to opioid overdose Coding of ADDE mortality data followed the methods outlined in Fuller et al.,55 based on the ICD-10 classification system. Specifically, we classified deaths as opioid overdose-related where codes for primary and/or secondary underlying causes of death included (see Table 3):
| Code | Cause of death description |
|---|---|
| F11–F19 | Mental and behavioural disorders due to psychoactive substance use |
| X40-X44 | Unintentional poisoning by and exposure to narcotics and psychodysleptics |
| X60-X69 | Intentional self-poisoning by and exposure to narcotics and psychodysleptics |
| X85 | Assault (homicide) by drugs, medicaments and biological substances |
| Y10-Y19 | Poisoning by and exposure to narcotics and psychodysleptics (undetermined intent) |
| T40 | Opium |
| T40.1 | Heroin |
| T40.2 | Other opioids (morphine, oxycodone, hydrocodone) |
| T40.3 | Methadone |
| T40.4 | Synthetic opioids excluding methadone (fentanyl, propoxyphene, meperidine) |
After linkage with the WDSD cohort, we identified three subcohorts: these comprise n = 1105 cases with deaths related to opioid overdose; a further n = 237,212 cases with deaths from all other causes; and remaining n = 2,989,079 individuals, still alive at their study end date.
Tables 4–6 summarise demographic data for these three study subcohorts.
| Gender | Study subcohort | |||||
|---|---|---|---|---|---|---|
| Alive at study end date (n = 2,989,079) | Death: all other causes (n = 237,212) | Death: opioid overdose-related (n = 1105) | ||||
| Male | 1,486,846 | 49.7% | 118,125 | 49.8% | 785 | 71.0% |
| Female | 1,502,211 | 50.3% | 119,087 | 50.2% | 320 | 29.0% |
| Missing | 22 | < 0.1% | 0 | 0 | ||
| Age band (years) | Study subcohort | |||||
|---|---|---|---|---|---|---|
| Alive at study end date (n = 2,989,079) | Death: all other causes (n = 237,212) | Death: opioid overdose related (n = 1105) | ||||
| 18–25 | 445,029 | 14.9% | 868 | 0.4% | 49 | 4.4% |
| 26–30 | 260,684 | 8.7% | 766 | 0.3% | 93 | 8.4% |
| 31–35 | 251,301 | 8.4% | 1046 | 0.4% | 152 | 13.8% |
| 36–40 | 231,821 | 7.8% | 1490 | 0.6% | 198 | 17.9% |
| 41–45 | 209,045 | 7.0% | 2257 | 1.0% | 142 | 12.9% |
| 46–55 | 459,262 | 15.4% | 9914 | 4.2% | 279 | 25.2% |
| 56–65 | 459,656 | 15.4% | 21,153 | 8.9% | 114 | 10.3% |
| 66–75 | 374,612 | 12.5% | 45,551 | 19.2% | 44 | 4.0% |
| 76 + | 297,669 | 10.0% | 154,167 | 65.0% | 34 | 3.1% |
| WIMD quintile | Study subcohort | |||||
|---|---|---|---|---|---|---|
| Alive at study end date (n = 2,989,079) | Death: all other causes (n = 237,212) | Death: opioid overdose-related (n = 1105) | ||||
| WIMD available | 2,863,661 | 231,794 | 1083 | |||
| 1 (most deprived) | 549,492 | 19.2% | 47,231 | 20.4% | 425 | 39.2% |
| 2 | 563,071 | 19.7% | 48,666 | 21.0% | 297 | 27.4% |
| 3 | 580,844 | 20.3% | 49,131 | 21.2% | 177 | 16.3% |
| 4 | 581,988 | 20.3% | 45,999 | 19.8% | 106 | 9.8% |
| 5 (least deprived) | 588,266 | 20.5% | 40,767 | 17.6% | 78 | 7.2% |
We next categorised records in EDDS, PEDW, CCD and SMDS for the 36 months up to immediately prior to and including an individual’s study end date, censoring at the start of the study window where necessary. For each attendance, admission or presentation, we used further routine data and appropriate coding framework (ICD-10: as per Table 2; codes 10B, 10C, 10D, 10Z and 31B in the NHS Wales Data Dictionary) to determine whether that event was related to opioid overdose, and further recording whether or not the event occurred within 1 month or within 12 months of an individual’s study end date.
Emergency department attendances
After linkage with the WDSD cohort, we identified EDDS records in n = 1,187,033 cases, with breakdown of the number of attendances by subcohort and period summarised in Table 7.
| Time period and attendance category | Number of attendances | Study subcohort | |||||
|---|---|---|---|---|---|---|---|
| Alive at study end date (n = 2,989,079) | Death: all other causes (n = 237,212) | Death: opioid overdose-related (n = 1105) | |||||
| Within 36 months | |||||||
| All attendances | 0 | 1,985,462 | 66.4% | 54,633 | 23.0% | 268 | 24.3% |
| 1–3 | 876,091 | 29.3% | 128,365 | 54.1% | 462 | 41.8% | |
| 4–12 | 121,955 | 4.1% | 51,322 | 21.6% | 306 | 27.7% | |
| 13 + | 5571 | 0.2% | 2892 | 1.2% | 69 | 6.2% | |
| Opioid overdose-related attendances | 0 | 2,975,604 | 99.5% | 235,316 | 99.2% | 887 | 80.3% |
| 1–3 | 13,054 | 0.4% | 1836 | 0.8% | 206 | 18.6% | |
| 4 + | 421 | < 0.1% | 60 | < 0.1% | 12 | 1.1% | |
| Within 12 months | |||||||
| All attendances | 0 | 2,520,674 | 84.3% | 76,000 | 32.0% | 423 | 38.3% |
| 1–3 | 443,898 | 14.9% | 137,261 | 57.9% | 502 | 45.4% | |
| 4–12 | 23,772 | 0.8% | 23,541 | 9.9% | 165 | 14.9% | |
| 13 + | 735 | < 0.1% | 410 | 0.2% | 15 | 1.4% | |
| Opioid overdose-related attendances | 0 | 2,984,543 | 99.8% | 236,191 | 99.6% | 961 | 87.0% |
| 1+ | 4536 | 0.1% | 1021 | 0.4% | 144 | 13.0% | |
| Within 1 month | |||||||
| All attendances | 0 | 2,926,591 | 97.9% | 14,5622 | 61.4% | 815 | 73.8% |
| 1–3 | 61,609 | 2.1% | 91,259 | 38.5% | 279 | 25.2% | |
| 4 + | 879 | < 0.1% | 331 | < 0.1% | 11 | 1.0% | |
| Opioid overdose-related attendances | 0 | 2,988,521 | ~100.0% | 236,974 | 99.9% | 1064 | 96.3% |
| 1 + | 558 | < 0.1% | 238 | 0.1% | 41 | 3.7% | |
Hospital admissions
After linkage with the WDSD cohort, we identified PEDW records in n = 998,563 cases, with breakdown of the number of attendances by subcohort and period summarised in Table 8.
| Time period and admission category | Number of attendances | Study subcohort | |||||
|---|---|---|---|---|---|---|---|
| Alive at study end date (n = 2,989,079) | Death: all other causes (n = 237,212) | Death: opioid overdose-related (n = 1105) | |||||
| Within 36 months | |||||||
| All admissions | 0 | 2,192,076 | 73.3% | 36,309 | 15.3% | 448 | 40.5% |
| 1–3 | 668,456 | 22.4% | 119,399 | 50.3% | 476 | 43.1% | |
| 4–12 | 110,284 | 3.7% | 64,795 | 27.3% | 158 | 14.3% | |
| 13 + | 18,263 | 0.6% | 16,709 | 7.0% | 23 | 2.1% | |
| Opioid overdose-related admissions | 0 | 2,987,393 | 99.9% | 236,842 | 99.8% | 1019 | 92.2% |
| 1–4 + | 1686 | 0.1% | 370 | 0.2% | 86 | 7.7% | |
| Within 12 months | |||||||
| All admissions | 0 | 2,616,002 | 87.5% | 52,332 | 22.1% | 622 | 56.3% |
| 1–3 | 336,773 | 11.3% | 137,906 | 58.1% | 418 | 37.8% | |
| 4 + | 36,304 | 1.2% | 46,974 | 19.8% | 65 | 5.9% | |
| Opioid overdose-related admissions | 0 | 2,988,572 | ~100.0% | 236,992 | 99.9% | 1049 | 94.9% |
| 1 + | 507 | < 0.1% | 220 | <0.1% | 56 | 5.1% | |
| Within 1 month | |||||||
| All admissions | 0 | 2,904,066 | 97.2% | 94,044 | 39.6% | 910 | 82.4% |
| 1+ | 85,013 | 2.8% | 143,168 | 60.4% | 195 | 17.6% | |
| Opioid overdose-related admissions | 0 | 2,989,013 | ~100.0% | 237,146 | ~100.0% | 1081 | 97.8% |
| 1 + | 66 | < 0.1% | 66 | < 0.1% | 24 | 2.2% | |
Critical care admissions
After linkage with the WDSD cohort, we identified CCDS records in n = 35,489 cases, with relatively few cases with multiple admissions within this number; approximately 90% had a single admission over 36 months, and only 2% had more than two such admissions in this period. The breakdown of the number of attendances by subcohort and period summarised in Table 9; no coding was available to categorise admissions as opioid overdose-related or otherwise.
| Time period | Number of attendances | Study subcohort | |||||
|---|---|---|---|---|---|---|---|
| Alive at study end date (n = 2,989,079) | Death: all other causes (n = 237,212) | Death: opioid overdose-related (n = 1105) | |||||
| Within 36 months | |||||||
| All admissions | 0 | 2,974,728 | 99.5% | 216,220 | 91.2% | 959 | 86.8% |
| 1 + | 14,351 | 0.5% | 20,992 | 8.8% | 146 | 13.2% | |
| Within 12 months | |||||||
| All admissions | 0 | 2,984,216 | 99.8% | 219,825 | 92.7% | 987 | 89.3% |
| 1 + | 4863 | 0.2% | 17,387 | 7.3% | 118 | 10.7% | |
| Within 1 month | |||||||
| All attendances | 0 | 2,988,554 | ~100.0% | 224,073 | 94.5% | 1025 | 92.8% |
| 1 + | 525 | < 0.1% | 13,139 | 5.5% | 80 | 7.2% | |
Substance misuse treatment presentations
After linkage with the WDSD cohort, we identified SMDS records in n = 32,774 cases, with relatively few cases with multiple presentations within this number; under 10% had more than three such presentations in this period. The breakdown of the number of attendances by sub-cohort and period summarised in Table 10.
| Time period and presentation category | Number of attendances | Study subcohort | |||||
|---|---|---|---|---|---|---|---|
| Alive at study end date (n = 2,989,079) | Death: all other causes (n = 237,212) | Death: opioid overdose-related (n = 1105) | |||||
| Within 36 months | |||||||
| All presentations | 0 | 2,959,595 | 99.0% | 234,283 | 98.8% | 744 | 67.3% |
| 1 + | 29,484 | 1.0% | 2929 | 1.2% | 361 | 32.7% | |
| Opioid overdose-related presentations | 0 | 2,974,825 | 99.5% | 236,576 | 99.7% | 851 | 77.0% |
| 1 + | 14,254 | 0.5% | 636 | 0.3% | 254 | 23.0% | |
| Within 12 months | |||||||
| All presentations | 0 | 2,976,864 | 99.6% | 235,559 | 99.3% | 862 | 78.0% |
| 1 + | 12,215 | 0.4% | 1653 | 0.7% | 243 | 22.0% | |
| Opioid overdose-related presentations | 0 | 2,983,446 | 99.8% | 236,867 | 99.9% | 940 | 85.1% |
| 1 + | 5633 | 0.2% | 345 | 0.1% | 165 | 14.9% | |
| Within 1 month | |||||||
| All presentations | 0 | 2,987,754 | ~100.0% | 236,979 | 99.9% | 1076 | 97.4% |
| 1 + | 1325 | < 0.1% | 233 | 0.1% | 29 | 2.6% | |
| Opioid overdose-related admissions | 0 | 2,988,468 | ~100.0% | 237,173 | ~100.0% | 1083 | 98.0% |
| 1 + | 611 | < 0.1% | 39 | < 0.1% | 22 | 2.0% | |
Proportions of individuals with no records in routine data in the 12 months prior to study end date
Based on linkages between the WDSD cohort and EDDS, PEDW, CCDS and SMDS records, we assessed the numbers of individuals within cohorts with no records in the 12 months prior to study end date; the breakdown of numbers and proportions by subcohort and period is summarised in Table 11.
| Time period (months prior to study end date) | Study subcohort | |||||
|---|---|---|---|---|---|---|
| Alive at study end date (n = 2,989,079) | Death: all other causes (n = 237,212) | Death: opioid overdose-related (n = 1105) | ||||
| 12 months | 2,316,451 | 77.5% | 39,236 | 16.5% | 307 | 27.8% |
| 1 month | 2,857,958 | 95.6% | 81,882 | 34.3% | 755 | 68.3% |
Discriminant function data analysis
For analysis, we combined the group of those alive at their study end date with the group of deaths from all other causes and contrasted this combined group with the (much) smaller group of decedents from opioid overdose. We examined the extent to which factors and covariates (recorded in routine data) (1) are associated with death from opioid overdose; and (2) enable us to identify these decedents, or a relatively small subset of an overall population that contains most or all of them.
Prediction via logistic regression
We fitted alternative versions of the logistic regression model, using (1) raw rather than banded counts of EDDS attendances and PEDW admissions, (2) all admissions and attendances rather than just those coded as opioid-related and (3) omitting WIMD quintile as a factor, thereby including data on all but 22 individuals in fitting the model.
Sensitivity and specificity
We obtain predicted probabilities by a logistic transformation of the linear predictor, including the constant term.
Logistic regression analysis
The appropriate methodology here is logistic regression; we illustrate its potential effectiveness using age and gender. Table 4 above shows that the decedents from opioid overdose are disproportionately male – 785 males and 320 females; compared with a near-equal split in both other groups in Table 4, and hence in the combined group. This observed difference in proportions is highly statistically significant.
Table 5 shows differences in the age profiles across the three original groups; for the combined group, as described above, the mean (standard deviation) age (in years) is 50.7 (20.6), while the corresponding value for the 1105 deaths from opioid overdose is 44.3 (13.3). Again, this observed difference is highly statistically significant.
For (1), therefore, the raw data indicate that both variables – separately and jointly – appear as statistically significant explanatory variables in logistic regression models for the binary outcome of death from opioid overdose (or not). Table 12 gives details on the fitted model, including age, gender, WIMD quintile; and the number of EDDS, PEDW, CCDS and SMDS attendances within 12 months of study end date.
| Variable | Coefficienta,b | OR | 95% CI | p- value |
|---|---|---|---|---|
| Constant | –8.145 | |||
| Femalec | –0.774 | 0.461 | (0.403 to 0.527) | < 0.001 |
| Age (years) | –0.013 | 0.987 | (0.984 to 0.990) | < 0.001 |
| WIMD d | ||||
| Quintile 1 | 1.439 | 4.218 | (3.302 to 5.388) | < 0.001 |
| Quintile 2 | 1.188 | 3.280 | (2.551 to 4.217) | < 0.001 |
| Quintile 3 | 0.696 | 2.006 | (1.533 to 2.623) | < 0.001 |
| Quintile 4 | 0.244 | 1.276 | (0.951 to 1.712) | 0.104 |
| EDDSe (opioid-related attendances) | 2.726 | 15.273 | (12.024 to 19.401) | < 0.001 |
| PEDWc (opioid-related admissions) | 2.309 | 10.064 | (6.793 to 14.908) | < 0.001 |
| CCDSe (all admissions) | 2.235 | 9.343 | (7.429 to 11.751) | < 0.001 |
| SMDSe (drug-related attendances) | 3.094 | 22.071 | (17.872 to 27.256) | < 0.001 |
We fitted alternative versions of this model, using (1) raw rather than banded counts of EDDS attendances and PEDW admissions, (2) all admissions and attendances rather than just those coded as opioid-related, and (3) omitting WIMD quintile as a factor, thereby including data on all but 22 individuals in fitting the model.
All fitted models had broadly similar characteristics to those summarised in Table 12; factors and covariates were (generally) highly statistically significant, and the predicted probabilities of death from opioid overdose are lower for females, reduce with age, and increase with WIMD quintile, and opioid-related attendances and admissions in EDDS, PEDW, CCDS and SMDS.
For (2), we consider the predictive ability of fitted logistic regression models: we obtain predicted probabilities by a logistic transformation of the linear predictor, including the constant term. For illustration, based on the fitted model above, a female aged 23, living within a LSOA categorised as the most deprived (WIMD Quintile 1), and with an opioid-related attendance at a substance misuse centre but no opioid-related ED attendances or hospital or critical care admissions, would have a predicted probability of 0.00916 of death from opioid overdose.
For the current routine data cohort, these predicted probabilities (of death from opioid overdose) range from –0.00003 to –0.96150, distributed with considerable skewness – 99% of individuals have a predicted probability < 0.001; 99.5% of individuals have a predicted probability < 0.0035.
We can now, in any cohort with the requisite individual-level data, take the high-risk population to include the set of individuals each with a predicted probability greater than some (arbitrary) threshold value. For instance, with a threshold value of 0.0004, applying this model across the entire cohort identifies 708 of the 1083 decedents – a true positive rate (sensitivity) of 65.4%; however, it also identifies 646,750 individuals with no death from an opioid overdose. This gives a total of 647,458 individuals to include in the high-risk population. Table 13 gives the full 2 × 2 classification for this threshold value and shows a true negative rate (specificity) of 2448685/3226269 or 79.1%.
| Predicted status | |||
|---|---|---|---|
| Opioid overdose death | Alive or non-opioid death | ||
| Recorded status | Opioid overdose death | 708 | 375 |
| Alive or non-opioid death | 646,750 | 2,448,685 | |
The sensitivity, specificity and potential contributions to the high-risk population for a range of threshold values are shown in Table 14, along with a summary of known outcomes.
| Threshold value | Opioid overdose decedents identified (sensitivity, %) | Identified for high-risk population | Total deaths in high-risk population | Excluded from high-risk population and without opioid overdose-related mortality (specificity, %) |
|---|---|---|---|---|
| 0.0002 | 933 (86.1) | 1,597,342 | 87,797 | 1,499,026 (48.4) |
| 0.0003 | 809 (74.7) | 989,151 | 60,980 | 2,107,093 (68.1) |
| 0.0004 | 708 (65.4) | 647,458 | 40,751 | 2,448,685 (79.1) |
| 0.0005 | 662 (61.1) | 432,962 | 25,943 | 2,663,135 (86.0) |
| 0.0006 | 566 (52.3) | 312,131 | 20,006 | 2,783,870 (89.9) |
| 0.0007 | 473 (43.7) | 193,751 | 17,597 | 2,902,157 (93.8) |
| 0.0008 | 381 (35.2) | 108,487 | 16,677 | 2,987,329 (96.5) |
| 0.0009 | 350 (32.3) | 57,022 | 15,839 | 3,038,763 (98.2) |
| 0.0010 | 346 (31.9) | 28,345 | 14,827 | 3,067,436 (99.1) |
Table 14, in conjunction with Tables 7 – 13, illustrates the extent of the ability of predictive methods using this set of routine outcomes; the majority of decedents from opioid overdose are essentially indistinguishable both from other decedents and those still alive at their study end date. As the latter two groups are considerably more numerous than the cohort of decedents from opioid overdose, it follows that one must either include in the high-risk population a substantial minority (or even a majority) of the overall population, or restrict the high-risk population to a relatively small proportion of the overall population, with a lower proportion of decedents from opioid overdose.
Accuracy of discriminant function
From the data we had available, we were not able to distinguish between decedents from opioid overdose and other decedents and those still alive using the discriminant function, partly due to the number of decedents from opioid overdose cohort being considerably smaller.
Based on logistic regression models we would need to monitor approximately one-third of the population to capture 75% of the decedents from opioid overdose in 1-year follow-up. Furthermore, as mortality in the 1-year period is estimated to be 6% of the third of the population, it is estimated that only 1–2% of these deaths would be categorised as a result of opioid overdose. As a high proportion of this population have no records of a healthcare event in the 12 months prior to death, it is generally the case that usage by decedents is matched or surpassed by individuals in other cohorts, therefore making the predictive link between death and healthcare events associated with opioid overdose weak.
Summary
We have provided details on healthcare utilisation and proportion of decedents from opioid overdose with no such records up to 12 months prior to death. We found a small number of decedents, which is a limiting factor in the development and accuracy of the discriminant function. Predictive variables produced in this study have shown that there are statistically significant differences in the factors and covariates in opioid overdose decedents and other cohorts, however, at an individual level and not sufficiently different from members of the other cohorts.
In order to identify and include 75% of the high-risk population for 1-year follow-up, we would need to monitor approximately one-third of the population. As only 1–2% of the deaths within the follow-up period would be the result of opioid overdose, it follows that reducing substantially or even eliminating entirely such deaths would have little overall effect on mortality rates in the high-risk population. Given the weak predictive link between deaths from and healthcare events related to opioid overdose, using routine data to identify with sufficient precision those at high risk of death from opioid overdose currently seems infeasible.