Key words
cancer - certification - survival analysis
Introduction
According to the National Cancer Plan, all cancer patients should receive treatment
in accordance with evidence-based treatment guidelines [1]. With this aim, the German Cancer
Society (Deutsche Krebsgesellschaft, DKG) has, as of 2003, established a
certification programme that focuses on structuring the entire process of care in an
evidence-based, guideline-adherent manner and is currently the largest in Europe
[2]
[3]. In order to obtain a DKG certificate,
hospitals need to meet a specified set of professional and quality requirements
based on S3-guidelines [2]. These
requirements cover the entire process of oncological care. Structural requirements
include e. g. multidisciplinary communication, psychooncological support, as
well as connection to the outpatient sector, social care, and rehabilitation.
Certified hospitals need to file annual reports via entity specific surveys and
indicator sheets covering key figures, part of which are quality indicators as
defined in S3-guidelines to retain the certificate. Requirements, surveys and
indicator sheets are publicly available via the DKG website.
Hospitals that do not hold a certificate may meet the same structural requirements
for cancer therapy, but are not obliged to do so. It is hence reasonable to assume
that the measures required to meet certification criteria ultimately improve
outcomes. The aim of the “WiZen”- study whose results serve as a
basis for this article was to provide reliable evidence about the effectiveness of
certification: The “WiZen Study” (Wirksamkeit der Versorgung in
onkologischen Zentren/Effectiveness of care in oncological centres), funded
by the Innovation Fund of the Joint Federal Committee (Gemeinsamer Bundesausschuss,
G‑BA, Funding number: 01VSF17020), provides a large and comprehensive analysis of
survival in hospitals certified by the DKG vs. non-certified hospitals on the basis
of nationwide AOK data and data from several clinical cancer registries. The study
finds that - irrespective of the entity - treatment in a certified hospital
increases the chances of survival of patients with incident cancer [4]
[5]
[6], consolidating previous evidence on
beneficial effects of certification both nationally [7]
[8]
[9]
[10] and internationally [11]
[12]
[13]
[14]
[15]. The statutory health insurance AOK
was covering, as of 2017, a total population of around 22 million adults. The study
contains cohorts of patients with incident cancer for 11 entities in total ranging
from 10,596 patients (cervical carcinoma) to 172,901 patients (lung carcinoma) in
the years 2009–2017. Survival analysis was conducted for eleven entities
separately, including Kaplan-Meier-estimates and Cox regression with shared frailty.
These entities were colon and rectal cancer, lung cancer, pancreatic cancer, breast
cancer, ovarian, endometrial and cervical cancer, prostate cancer, head and neck
cancers and brain tumors, defined via ICD-10 codes, see Supporting [Table 1]. For each entitity, a set of
covariates was considered that consisted of patients’ demographic
information (age, sex), disease-related information (distant metastasis, secondary
malignoma, comorbidities) and hospital-level information (hospital status –
teaching, university hospital, and ownership, as well as number of beds) and the
calendar year of treatment to take into account effects of medical progress. The
relative survival advantages were between 3 and 26 percent for the 11 entities and
cohorts studied ([Fig. 1]).
Fig. 1 Adjusted hazard ratios (95% confidence intervals) of
overall survival for treatment in DKG-certified vs. non-certified hospitals
for the eleven entifies considered.
Table 1 Analysis of Life Years Lost per year for the tumour
entities investigated.
|
Entity
|
n (non-certified)
|
proportion (%) (non-certified)
|
Difference in area between survival functions
|
Difference for non- certified population (YLL)
|
Potentially lost life years YLL/yr in Germany (year of
reference, 2017)
|
|
Colon cancer
|
68,826
|
62.7
|
0.21
|
14,495
|
5,114
|
|
Rectal cancer
|
29,370
|
57.1
|
0.24
|
7,042
|
2,484
|
|
Pancreatic cancer
|
39,892
|
88.0
|
0.17
|
6,649
|
2,346
|
|
Breast cancer
|
52,451
|
36.5
|
0.29
|
15,465
|
5,456
|
|
Cervical cancer
|
16,031
|
77.1
|
0.32
|
5,078
|
1,791
|
|
Endometrial cancer
|
7,769
|
73.3
|
0.38
|
2,984
|
1,053
|
|
Ovarian cancer
|
24,222
|
80.5
|
0.13
|
3,116
|
1,099
|
|
Lung cancer
|
139,115
|
80.0
|
0.05
|
7,152
|
2,523
|
|
Prostate cancer
|
57,112
|
70.0
|
0.25
|
14,305
|
5,047
|
|
Brain tumors
|
58,032
|
92.5
|
0.19
|
11,304
|
3,988
|
|
Head and neck cancers
|
44,576
|
84.5
|
0.15
|
6,642
|
2,343
|
|
total
|
537,396
|
-
|
|
-
|
33,243
|
Across entities, less than half of patients with incident cancer were treated in
certified hospitals during the study period (2009–2017) and the proportion
of patients treated in certified hospitals was 31.3% during the observation
period. The proportion of patients who have been not been treated in a certified
hospital ranges from 36.5 (breast cancers) to 92.5 per cent (brain cancers), [Fig. 2].
Fig 2 Proportion of patients that have/have not received treatment in
a DKG-certified hospital per entity, along with the number of patients in
each entity.
Given that, as a result of the WiZen study, the certification effect showed a benefit
in survival across entities, the question arises what this benefit would have
encompassed had all patients been treated in a certified center. While survival
analysis constitutes a powerful tool to evaluate effect of treatments for diseases
such as cancer its direct results are not well suited to quantify this benefit:
Cox regression addresses the effect of multiple variables upon the survival time, and
is hence suited to quantify the effect of certification. However, results of a Cox
regression are typically presented in terms of Hazard ratios, i. e. the
relative impact a given variable has on the (time-dependent) hazard with respect to
a given reference level, e. g. presence vs. absence of a covariate. The
hazard ratio constitutes a relative measure whereas in many settings the impact on a
given population in terms of absolute numbers is required. This is particularly true
for patients, who can interpret an absolute risk or chance such as the number needed
to treat (NNT) much better than a relative risk or hazard ratio. For health policy
makers, quantification of the total absolute effect as characterized by Years Life
Lost (YLL) is of high importance [16]. In
addition, the interpretation of a hazard ratio requires a specialized background in
survival analysis and calls for careful communication [17]. The hazard ratio is thus critisized
for not being particularly suited to illustrate study findings to, e. g.
decision makers and other stakeholders in health care and concepts to avoid using it
altogether are emerging [18]
[19]
[20]
[21].
In this article, we compute two absolute measures from adjusted survival curves that
incorporate the results from Cox regression, but do not rely on solely reporting the
hazard ratios. The first measure is the number needed to treat (NNT), which allows
to quantify the difference in deaths after a given point in time after diagnosis
between two groups. The second measure is years of life lost (YLL), which is
commonly used in the communication of statistical assessment for burden of disease
for e. g. cancer and diabetes by. e. g. the Robert-Koch-Institute
(RKI) and the National Cancer Institute (NCI) [22]
[23]
[24]. We deduct both measures from
survival curves as predicted by adjusted Cox regression, allowing us to focus on the
cerfication effect only.
Our approach allows for combination of results from multivariate statistical analysis
with a framework that we feel is suited for communication of the effectiveness of a
health-care intervention with non-statisticians.
Methods
We use Cox regression with shared frailty [25]
[26] for each entity separately for a
given set of covariates detailed in [4]
and summarized in Table S2. The resulting hazard ratios for each covariate indicate
the extent to which the prognosis changes relatively with respect to a reference
level of each variable. A hazard ratio>1 indicates a poorer prognosis for
the assoated variable compared to the reference level, and a hazard ratio<1
indicates a beneficial effect. The Cox regression assumes that on a baseline which
is determined in the regression, the adjusted hazard behaves proportionally to the
baseline hazard with respect to the hazard ratio of a specific covariate
(“proportional hazards assumption”). The probability of survival (or
death) is then calculated as a prediction of a survivor function, based on an
exponential transformation of the adjusted hazard function . This prediction of a
survivor curve, is based on the entire model result, i. e. the hazard
(ratios) of all covariates as well as the baseline hazard. The prediction must be
made on a model population. The model population used here is defined by the mean
value over all covariates of the actual population, i. e. mean value of age
group, mean value of sex m/f, mean value of oncological second disease, etc.
Stratifying variables are excluded from mean calculation. The prediction based on
the model results on the model population yields an adjusted survival curve. This
survival function in analogy to a Kaplan-Meier-curve starts with value one at time
zero and decreases over time in a non-linear fashion. Since we are interested in a
survival probability in certified/non-certified centres, the survival
function is computed upon stratification by certified/non-certified
hospitals.
Hence, two predictions using all estimators from the model on the mean model
populations are performed, for an “all certified” and an
“all non-certified” case. The covariates, with the exception of
certification, are therefore identical for both strata. Fig. 3 shows a schematic
representation of these two predictions. From these predictions, we compute two
measures, i) the number of life years lost and ii) the number needed to treat (NNT)
for one additional patient to survive at least five years after diagnosis. As the
survival function is time-dependent, all considerations that result in absolute
measures must be supplemented with a time reference.
Life Years Lost (YLL)
The years of life lost due to initial treatment in a certified center compared to
a non-certified hospital correspond to the area separated by the two survival
curves. It is therefore the difference between the areas under the respective
survival curves. As the time of follow-up is restricted, we introduce a cutoff
to the area, which corresponds to censoring for all values that exceed cutoff.
Our total observation time is nine years; we set the cutoff to eight years to
account for uncertainty in the prediction towards the end of the observation
time, hence estimating a lower bound of the area. This area is now rescaled with
the population in the non-certified setting, pop
ncert
, resulting in an estimate of Years of life lost (YLL) due to the fact
the hospitals were non-certified:
Note that in our variant for life years lost, we do not explicitely take into
account a person’s age at onset of disease as is common in epidemiology
for computation of years of life lost due to mortality. Here, the age at the
onset of diagnosis is incorporated as covariate into the prediction of the
survival curve and through the median population.
NNT and avoidable deaths within 5 years of diagnosis
The number needed to treat (NNT) related to the certification effect, is given by
the inverse difference in the two survival functions for a given time t
surv
. We set this survival time to t
surv
.=5 years, as the 5-year survival is an important outcome and a
broadly used epidemiological measure for the burden of disease in oncology. This
time period is also relevant for patients because after this time the incidence
of recurrences is significantly reduced in most cases and follow-up care is
usually also terminated.
where s(t) is
the simulated survival at time t for stratum X. Rescaling with the
population in the non-certified setting results in an estimate of the number of
deaths that could have been avoided within 5 years from diagnosis:
Results
Adjusted survival functions were computed based on Cox regression with shared frailty
for each entity separately for a given set of covariates. This set was identical for
each entity with the exception of comorbidities that are entity-specific as defined
by clinical experts, see Supporting Table S2. Within the WiZen study, we fitted Cox
regression models upon gradually increasing the sets of covariates and found that
the cerfication effect does not depend substantially on the choice of model [4]. The concordance (Harrels’ C)
for these models is increasing with model complexity. The increase is substantial
upon adding disease-related information to the core set (certification, age, sex),
and marginal upon addition of hospital-related covariates and year of diagnosis. It
becomes maximal and ranges from 0.67 to 0.82 for the model including the full set of
covariates across entities, Supporting Table S3. We thus compute adjusted survivals
for each entity from the model with the best concordance, i. e. the one
including all covariates, with hazard ratios ranging from 0.77 to 0.92 across
entities, Supporting Table S4.
Life Years Lost (YLL)
The estimation of the potential of care through treatment of patients with
incident cancer into certified hospitals is based on the difference of the area
under the adjusted survival function of the treatment in certified hospitals and
the treatment in non-certified hospitals simulated from the Cox regression (see
Fig. 3). Due to the limited observation period of the WiZen study of 9 years,
the period up to 8 years after diagnosis was considered as the cutoff limit. The
results are therefore conservative and include only the years of life lost
within this period.
Taking into account the proportion of the national population insured by AOK,
which was AOK
coverage 2017
=31.5% (people insured by the AOK as of July 2017:
25.990.759, German population 2017: 82.522.000, [27]), we estimate for “Life
Years Lost/yr in Germany” based on the WiZen project results for
the overall population in Germany from:
The following [Table 1] shows the Life
Years Lost for the 11 entities examined. The number of people affected for each
entity determines the benefit in survival on a population level. The total
number of patients treated in a non-certified hospital ranges from 7,769 for
endometrial to 139,115 patients for lung cancer. The size of the population
hence adds an essential contribution to the total YLL. The size of the area
between survival functions serves as a guideline to the extent of the
certification effect for each entity: if the area increases the benefit in
overall survival increases as well. Hence, the YLL becomes maximal for entities
that have a larger benefit, but also many people affected, as is the case for
e. g. breast, colon and prostate cancer in contrast to lung cancer
(small effect) or endometrial cancer (less people affected).
In total, there is a potential of around 33,200 life years saved per year in
Germany.
NNT and avoidable deaths within 5 years of diagnosis
In a next step, the number needed to treat is computed based on the WiZen
results. Using the population of patients treated in certified and non-certified
centres, we then use the NNT which indicates the potential to avoid one death
within 5 years of diagnosis to assess the potential for the total of avoided
deaths within 5 years after diagnosis.
The calculation of avoidable deaths results from the difference of the survival
curves after 5 years (Fig. 3), in analogy to the procedure for Life Years Lost
related to the federal population 2017 as:
The results are shown in [Table 2] for
each entity individually. For cancers with very low chance of (crude) survival
after five years, such as pancreatic cancer and lung cancer, the difference in
the survival function is small as well, reflecting the lesser (overall) chance
of preventing death with fighting the most deadly cancers. In analogy to YLL,
the number of avoidable deaths scales with the size of the population for each
entity, and, as a consequence, a large number of people who have not been
treated in a certified setting provides a larger potential for avoidable deaths
5 years post-diagosis, which we find for colon, breast, and prostate cancer. In
total, approx. 4,700 deaths per year could have been avoided 5 years post
diagnosis if all patients had been treated in a certified cancer.
Table 2 Analysis of potentially avoidable deaths within 5
years after diagnosis per year for the tumour entities
investigated.
|
Entity
|
n (non- certified)
|
proportion (%) (non- certified)
|
crude 5-yr survival rate, non-certified
|
difference certified/non-certified from adjusted
5-yr-survival rate*
|
number needed to Treat (NNT)*
|
deaths 5 yrs post-diagnosis that could have been
avoided/yr (Germany)*
|
|
Colon cancer
|
68,826
|
62.7
|
0.467
|
0.031
|
32
|
754
|
|
Rectal cancer
|
29,370
|
57.1
|
0.433
|
0.036
|
28
|
372
|
|
Pancreatic cancer
|
39,892
|
88.0
|
0.065
|
0.014
|
70
|
202
|
|
Breast cancer
|
52,451
|
36.5
|
0.719
|
0.046
|
22
|
859
|
|
Cervical cancer
|
16,031
|
77.1
|
0.357
|
0.046
|
22
|
258
|
|
Endometrial cancer
|
7,769
|
73.3
|
0.533
|
0.057
|
18
|
156
|
|
Ovarian cancer
|
24,222
|
80.5
|
0.650
|
0.020
|
51
|
168
|
|
Lung cancer
|
139,115
|
80.0
|
0.169
|
0.006
|
179
|
274
|
|
Prostate cancer
|
57,112
|
70.0
|
0.712
|
0.039
|
26
|
789
|
|
Brain tumors
|
58,032
|
92.5
|
0.480
|
0.027
|
37
|
555
|
|
Head and neck cancers
|
44,576
|
84.5
|
0.453
|
0.022
|
46
|
341
|
|
total
|
537,396
|
-
|
-
|
-
|
-
|
4,729
|
* from adjusted survival curve.
Discussion
We illustrated how to derive Life years lost and the Number needed to treat from
adjusted survival functions that were computed based on the results of the WiZen
study. We have derived these measures from the cohort used in the WiZen study which
includes AOK-insured patients that were diagnosed with one out of eleven types on
cancers within the years 2009–2017. Based on the 537 396 patients or
68,7% of the study population in the cohort who have not received treatment
in a certified hospital, we estimated a total of 33 243 YLL per year for the entire
German population as of 2017. The corresponding potential to avoid death cases for
as long as five years within diagnosis sums up to 4 729 per year in Germany. Both
YLL and NNT depend on the entity specific survival that was estimated based on the
Cox regressions, as well as the size of the population under consideration. The
difference in survival curves (both in area and for the 5 year limit) tends to be
smaller for cancers with overall low survival prospects, such as pancreatic cancer.
Beyond this observation, we currently do not have any additional information about
the range of differences, which may arise from many factors both on the cohort and
on the intervention level. As the size of the population under consideration varies
significantly, the greatest contribution to the total deaths that could be avoided
arise from cancers that have either high incidence such as e. g. breast and
colon cancer or a very low proportion of certified centers such as brain tumours. To
date, we are not aware of any literature about YLL and the assesment of avoidable
deaths in the context of certification, beyond the national [7]
[8]
[9]
[10] and international [11]
[12]
[13]
[14]
[15] evidence about the benefits of
structuring the process of cancer care, be it via certification or accredition.
Our findings have a set of limitations: As in any communication of statistical
results, it needs to be pointed out that the estimates presented here are based on
various assumptions and do, to some extent, depend on the method [28]. The rescaling from the original
population that covers patients insured by the AOK,which is roughly one third of the
German population, to the entire size of the population serves as a rough estimate
to estimate the total YLL and the number of deaths within 5 yrs from
diagnosis that could have been avoided. The validity of this extrapolation is based
on the assumption that the cohort on which the survivals were predicted adequately
represents the epidemiology of the disease in Germany, which we feel is a valid
assumption given the large size of the cohort. Another limitation and one of the
most unfortunate shortcomings of health insurance data is that these contain very
litte information on important cancer-related measures such as staging and grading.
The quality of the data, and hence, the model prediction could be improved by
e. g. linking SHI data with data from cancer registries as evaluated in
[29].
The strength of this assesment lies in that our estimates avoid the reporting, and
thus potential for misinterpretation, of the hazard ratio and can be used to inform
political decision makers about the extent of the benefit in survival found in the
WiZen study. As our estimates provide absolute results such as the amount of deaths
within five years of diagnosis, they can be used in health-economic analyses that
deal with e. g. certification and cost-effectiveness [30].
Fig 3 Schematic representation of the calculation of Life Years Lost, corresponding to the area between two adjusted survival curves in the period
up to 8 years ("cutoff"). Also shown is the difference in survival after 5 years, as used in the calculation of avoidable deaths.
Conclusion
For each individual entity, steering into certified centres would have a relevant
effect on preventable deaths within 5 years of diagnosis. The Number Needed To
Treat, i. e. the number of additional patients to be treated in a certified
centre in order to avoid a death 5 years post-diagosis, also depends on the general
prognosis of the entities. For example, the Number Needed To Treat is higher for
cancers with a generally poor prognosis, such as lung cancer or pancreatic
cancer.
Through illustrating the potential in survival benefit from certification in the past
decade, our analysis provides a starting point for a broader discussion of political
implications that would either foster certification and/or install a
coordinated effort to steer patients into certified hospitals.
Ethics declaration
The WiZen study was approved by the ethics committee of the TU Dresden (approval
number: EK95022019, IRB 00001473, OHRP IORG0001076). Data processing and analyses
was conducted in line with the Declaration of Helsinki and the General Data
Protection Regulation of the European Union.
Contributor’s Statement
None
