Towards better guidance on caseload thresholds to promote positive tuberculosis treatment outcomes: a cohort study

TB cases

A total of 67,869 TB cases not diagnosed post-mortem, with treatment expected to last
12 months or less, were notified in England between 2003 and 2012 (Additional file
2). Caseload was calculated as a mean for each year from the preceding 3 years of data,
thus cases notified from 2003 to 2005 were utilised to assess caseload for 2006 and
could not be included in the analysis. Of the remaining 48,838 cases, 1,468 were missing
a treatment outcome, 7,943 a designated treating clinician, and 1,336 a named treating
hospital. Table 1 describes the baseline characteristics of these cases, who were generally male, aged
20–39 years of age, had not been previously diagnosed with TB, and had no social risk
factors. As documented in previous reports, England has a high proportion of extrapulmonary
TB cases compared to some other low-incidence settings 1], 16]. Of all cases, only 3.6 % were resistant to EMB, INH, or PZA (6.1 % of those with
a culture result).

Table 1. Univariate logistic regression of the association between clinician caseload and treatment
outcome

Clinician caseload

Overall, 2,276 unique clinicians were present using the main groupings and 2,351 with
the alternate groupings. Using the former, there was a median of 6 clinicians per
hospital across the entire time period (interquartile range, 2–12) and a median caseload
of 15 TB cases per clinician over the preceding 3 years (interquartile range, 4–42).
Little change in the percentage of cases being managed by clinicians below the toolkit
threshold was seen from 2006 to 2012 (42.4 % in 2006, 38.7 % in 2012; Fig. 1a), although calculating caseload for the previous year only (allowing examination
of caseload additionally in 2004 and 2005) indicated a drop in the percentage between
2004 and 2006 (51.0 % in 2004, 36.8 % in 2006). Additionally, little redistribution
was seen from very low caseloads (less than 2 or between 2 and 5) to higher ones over
time (data not shown).

Fig. 1. Mean clinician and hospital tuberculosis patient caseloads over the preceding 3 years,
England, 2006–2012. Percentage of tuberculosis cases a managed by a clinician who saw fewer than 10 cases on average over the preceding
3 years, b treated by a hospital with differential quartiles of caseload over the preceding
3 years. Dotted line year toolkit released. Error bars are 95 % confidence intervals

Clinician caseload and TB case outcomes

Of the 48,838 cases within this cohort, 1,468 (3.0 %) were missing outcome information,
6,777 (13.9 %) had an unfavourable outcome, 942 (1.9 %) a neutral outcome, and the
remainder a good outcome (Table 1). Of individuals with an unfavourable outcome, 940 (13.9 %) had died due to the factors
listed in Additional file 1, 2,229 (32.9 %) had been lost to follow-up, 3,104 (45.8 %) were still on treatment,
and 504 (7.4 %) had had their treatment stopped. Of cases managed by clinicians below
the toolkit threshold 15.0 % had an unfavourable outcome versus 12.7 % of cases managed
by clinicians above it; 15.6 % of cases in the pre-toolkit period had an unfavourable
outcome versus 13.5 % post-toolkit.

There was very strong evidence for clinician caseload being associated with the odds
of having an unfavourable outcome in a univariate model; cases managed by clinicians
below the threshold had higher cluster-specific odds than those managed by clinicians
above it (cluster-specific odds ratio (OR), 1.20; 95 % CI, 1.10–1.30; P 0.001; Table 1). Small amounts of within-clinician correlation were observed with very strong evidence
(??=?0.05; 95 % CI, 0.04–0.07; P 0.001); the large number of cases within some clinician groups meant it was necessary
to confirm the reliability of each model’s parameter estimate approximations using
different numbers of cut points. All potential confounders had evidence of a strong
association with the outcome (Table 1). Being female, an ethnicity other than White, or notified within London were associated
with lower cluster-specific odds of an unfavourable outcome. Being notified pre-toolkit
was associated with higher cluster-specific odds of an unfavourable outcome (likely
due to general improvements in care over time), as was shared clinical management.

All potential confounders, aside from drug sensitivity (which was not associated with
clinician caseload in a univariate model), were considered for inclusion in the final
model using a backwards deletion strategy. At the end of this process, eight remained:
notification date, location, sex, age, ethnic group, having previously had TB, social
risk factors, and shared clinical management. Including age as a linear variable did
not improve model fit. There was very limited statistical evidence for effect modification.
In the final multivariable model, ? was similar to that of the univariate model (0.04;
95 % CI, 0.03–0.05; P 0.001) and estimate approximations were again found to be reliable.

The multivariable model also demonstrated very strong evidence for higher cluster-specific
odds of an unfavourable outcome being associated with management by clinicians below
the caseload threshold (cluster-specific OR, 1.14; 95 % CI, 1.05–1.25; P?=?0.002; Table 2). There was weak evidence for an association between treatment outcome and shared
clinical management (cluster-specific OR, 1.37; 95 % CI, 0.97–1.94; P?=?0.08).

Table 2. Multivariable logistic regression of the association between clinician caseload and
treatment outcome

Hospital caseload

Although the Department of Health’s toolkit suggests a threshold at the level of the
clinician, it may be more appropriate to recommend one at the level of the hospital
due to the relative ease of calculation and its ability to better capture the functionality
of an entire case management team. Hospital caseload was divided into quartiles: 27,
27–72, 73–113, 114. These figures were conveniently similar to the thresholds proposed
by Cegolon et al. 6] and Lake et al. 7]. When hospital caseload was averaged over the last 3 years, little change in the
proportion of cases being managed below the toolkit threshold was seen from 2006 to
2012 (Fig. 1b).

Hospital caseload and TB case outcomes

Overall, 15.9 % of cases treated in a hospital with less than the lower quartile of
caseload had an unfavourable outcome versus 12.8–13.4 % in hospitals above it (Table 3). There was very strong evidence for hospital caseload being associated with the
cluster-specific odds of having an unfavourable outcome in a univariate model (P 0.001; Table 4). With the upper quartile of caseload as the baseline, only the CI of the smallest
caseload strata did not cross the null (cluster-specific OR, 1.32; 95 % CI, 1.13–1.55).
In this model, there was very strong evidence for weak within-hospital clustering
(??=?0.03; 95 % CI, 0.02–0.04; P 0.001). All potential confounders had evidence of a very strong association with
the outcome (Table 4).

Table 3. Hospital caseload and treatment outcomes

Table 4. Univariate and multivariable logistic regression of the association between hospital
caseload and treatment outcome

Following the same process as for clinician caseload, a final multivariable model
was built for hospital caseload that retained notification date, PHE centre of notification,
sex, age, ethnic group, UK born, and social risk factors (Table 4). Interactions were detected between ethnic group (P?=?0.004) and being UK born (P 0.001) and hospital caseload. As ethnic group and being UK born had 10 possible
strata combinations between them, they were condensed for the final model into: White
UK born, Other UK born, White not UK born, Black not UK born, Indian subcontinent
not UK born, Other not UK born. The observed interaction was retained using this categorisation
(P?=?0.01). Among those not born in the UK, the likelihood of an unfavourable outcome
may increase with decreased caseload, although CIs generally crossed the null (Table 5). In the UK born, if anything, this pattern was reversed. The association between
the combined ethnic group and place of birth variable with treatment outcomes is presented
in Table 6; a clear pattern of effect was difficult to discern. In the final model there was
very strong evidence for weak within-hospital clustering (??=?0.03; 95 % CI, 0.02–0.04;
P 0.001).

Table 5. Multivariable logistic regression of the hospital caseload-treatment outcome association,
stratified by country of birth/ethnicity

Table 6. Multivariable logistic regression of the birth country/ethnicity-treatment outcome
association, stratified by hospital caseload

Sensitivity analysis

The following sensitivity analyses were undertaken: (1) all records with missing outcome
information were recoded as having a good/neutral outcome or an unfavourable outcome,
(2) all records with missing clinician caseload information were coded to be either
above or below the caseload threshold, (3) where present, the alternative named clinician
was used to calculate whether a case was managed by an individual above or below the
threshold. Analysis (2) was not deemed necessary for the hospital caseload model due
to relatively low levels of missingness. Effect estimates were relatively uniform
throughout (data not shown), which was especially reassuring given the apparent association
between hospital caseload and treatment outcome reporting (Table 3). Exclusion of INH resistant cases also did not appreciably alter effect estimates,
nor did the exclusion of patients with neutral treatment outcomes (data not shown).

Encouragingly, an analysis restricted to the 10 hospitals in which we undertook additional
clinician name and coding data checks to identify infectious disease or respiratory
consultants with greater certainty indicated that the threshold caseload of 10 was
still associated with treatment outcome, albeit with low power to test the hypothesis
of no association due to the radically reduced size of the dataset (cluster-specific
OR, 1.11; 95 % CI, 0.84–1.46; P?=?0.47; Additional file 3).

Utilising different clustering structures had little impact on the results observed
when point estimates and CIs were compared like-with-like in terms of the number of
included patients (Additional file 4).

In the clinician caseload model, the inclusion of hospital caseload did little to
alter the estimated association between clinician caseload and treatment outcome (cluster-specific
OR, 1.10; 95 % CI, 1.01–1.20; P?=?0.03 with clustering structure B and cluster-specific OR, 1.10; 95 % CI, 1.01–1.21;
P?=?0.03 with structure C). In the hospital caseload model, the inclusion of clinician
caseload generally reduced the point estimates observed with both clustering structures
(Additional file 5).

Different thresholds for clinician caseload

In order to explore the relationship post-toolkit between clinician caseload and treatment
outcomes at different caseload thresholds, two versions of the final multivariable
model were run. The first reduced the caseload threshold to one, as per Gardam et
al.’s findings 5], in order to explore the impact of very low average caseloads as might be seen in
low-incidence settings. The effect estimate seen was similar to those produced by
the 10 case threshold: (cluster-specific OR, 1.13; 95 % CI, 1.01–1.27; P?=?0.04; Additional file 6a). Recursive partitioning suggested that the caseload threshold creating the strongest
association between clinician caseload and treatment outcome was 12.666 cases per
year, which was associated with an adjusted OR of 1.15 (95 % CI, 1.04–1.27; P?=?0.01) post-toolkit (Additional file 6b).