Changing causes of death in persons with haematological cancers 1975–2016

Causes of death in persons with haematological cancers include the index cancer, a new cancer or a seemingly unrelated cause such as cardio-vascular disease. These causes are complex and sometimes confounded. We analyzed trends in cause of death in 683,333 persons with an index haematological cancer diagnosed in 1975–2016 reported in the Surveillance, Epidemiology and End Results dataset. Non-cancer deaths were described using standardized mortality ratios. The index cancer was the predominant cause of death amongst persons with plasma cell myeloma, acute lymphoblastic leukaemia and acute myeloid leukaemia. Non-cancer death was the major cause of death in persons with chronic lymphocytic leukaemia, Hodgkin lymphoma and chronic myeloid leukaemia, mostly from cardio-vascular diseases. The greatest relative decrease in index-cancer deaths was amongst persons with Hodgkin lymphoma, chronic myeloid leukaemia and chronic lymphocytic leukaemia, where the proportion of non-cancer deaths increased substantially. Changing distribution of causes of death across haematological cancers reflects substantial progress in some cancers and suggests strategies to improve the survival of persons with haematological cancers in the future.


INTRODUCTION
Haematological cancers are an important cause of death worldwide [1]. In 2020, there were an estimated 1.3 million new haematological cancers with 700,000 deaths worldwide. Incidences of most haematological cancers increase with age. The Global Burden of Disease (GBD) study reported that incidences of non-Hodgkin lymphoma (NHL) and leukaemias increased 45 percent and 26 percent from 2006 to 2016 [2].
There are substantial recent advances in therapy of some haematological cancers such as Hodgkin lymphoma (HL) and NHL, chronic myeloid and lymphocytic leukaemias (CML and CLL) and plasma cell myeloma (PCM) [3][4][5][6][7]. There are more modest recent advances in other haematological cancers such as acute myeloid and lymphoid leukaemias (AML and ALL). Much of this progress has resulted in increased survival rather than increased cures [8]. Survivors of cancer are at risk for a range of late-occurring side effects from treatment, which would also change the fate of these patients [9]. In older persons with a haematological cancer, increased survival after a cancer diagnosis may result in more deaths from age-related diseases such as cardio-vascular disease (CVD), stroke and diabetes.
Analyses of causes of death in persons with a haematological cancer are complex. For example, as deaths from some cancers such as HL have decreased the proportion of deaths from new cancers, such as therapy-related AML and CVD, have increased [10,11]. In CML, where survival has increased substantially, the proportion of deaths from non-CML causes has increased markedly [12]. Sometimes there is uncertainty to identify the cause(s) of death in persons with a haematologic cancer. For example, some persons with AML receiving a haematopoietic cell transplant whilst in 1 st remission die from transplant-related mortality, such as graft-versus-host disease (GvHD) [13]. If these persons were cured before they received a transplant, it is unknown whether the death should be scored as from AML, the transplant or GvHD.
We interrogated data from the Surveillance, Epidemiology and End Results (SEER) dataset 1975-2016 to assess the trends for cancer-related and 26 major types of non-cancer cause(s) of death of persons with a haematological cancer.

SUBJECTS AND METHODS Data source and study population
We performed a population-based cohort study using data from the Surveillance, Epidemiology and End Results (SEER) programme, a population-based cancer registry with data on cancer demographics, incidence, anatomic site, histology, stage, therapy, socio-economic state and vital statistics of about one-third of persons with cancer in the US [14].
We extracted data on all first primary newly-diagnosed haematologic cancers diagnosed between 1975 and 2016 from the SEER 18 database (2019 Edition) using SEER*Stat software version 8.3.8 [15]. Diagnosis was based on coding in International Classification of Diseases for Oncology 3rd edition [ICD-O-3] codes, 9590-9992. Subjects were excluded if their diagnoses were obtained only from death certificates or autopsies. We also excluded subjects without active follow-up and those with unknown follow-up time, age at diagnosis and/or cause of death (Fig. S1). To analyze the causes of death as a function of calendar year of death, we extracted cause of death data amongst persons with a first haematological cancer diagnosis between 1975 and 2016 using the SEER-9 incidence-based mortality session [16]. For comparison with the general population, we extracted sex-, age-, race-and year-stratified mortality data of the US general population between 1975 and 2016 from the National Center for Health Statistics [17].
The Sun Yat-sen University Cancer Center (SYSUCC) Institutional Review Board (IRB) waived the requirement for a Research Data Agreement and informed consent consistent with principles of the Declaration of Helsinki.

Co-variates analyzed
Follow-up from diagnosis was defined as the interval between cancer diagnosis and death from any cause, last follow-up or the end of the study on December 31, 2016. Interrogated co-variates included sex, race, age at diagnosis, year of cancer diagnosis, cancer histology, follow-up duration and cause of death.
We classified haematological cancers by histology including: (1) [ICD]. Because death codes changed over the interval we studied, we used ICD-8 codes for cases diagnosed from 1975 to 1978, ICD-9 codes for cases diagnosed from 1979 to 1998 and ICD-10 codes for cases diagnosed from 1999 to 2016. The SEER programme summarized these codes and generated a combined death code.
We classified causes of death into three groups: (1) index-cancer; (2) non-index-cancer; or (3) non-cancer. Causes of death were defined by SEER cause-specific death classification variable from death certificates [18,19]. Non-cancer causes were categorized into 26 major groups which were consolidated into 7 broad categories: (1) infection; (2) CVD; (3) respiratory disease; (4) gastro-intestinal disease; (5) kidney disease; (6) external injuries; and (7) other. Deaths from in situ, benign or unknown behavior neoplasms are classified as non-cancer deaths by SEER programme [18,19]. We did not include these deaths in the non-cancer deaths. These deaths also cannot be classified into cancer-related deaths; thus, they were excluded from our analyses (3,534 cases were excluded [0.5%]; Fig. S1). In the SEER programme, survival is measured in months. Subjects with survival < 1 month were recorded as 0 months according to standard epidemiologic convention [20].   The bold values are the indicate major groups.

Statistical analyses
Mortality rates were calculated as number of deaths divided by the personyears of follow-up. For non-cancer deaths, standardized mortality ratios (SMRs) and 95% Confidence Intervals (CIs) were calculated for comparison with the US population [19,21]. SMR was estimated as ratio of observed to expected number of deaths. Observed number of deaths represents the number of deaths from certain causes in persons with cancer and expected number of deaths represents numbers of people dying from the same causes in the general population with a similar distribution of age, sex, race and calendar year. All-cause and cause-specific SMRs were calculated among patients with haematological cancer.
To obtain the expected number of deaths, we derived the stratumspecific mortality rates from the same cause in the reference general population collected by the SEER programme and calculated person-years of relevant strata in the haematological cancer cohort. The stratum-specific expected number of deaths was estimated as the product of death rate in the reference cohort and person-years in the haematological cancer cohort. Total expected number of deaths is the sum of all expected deaths across the strata.
Age at diagnosis was divided into 19 cohorts: (1) 00 years; (2)  confidence intervals (CIs) of SMRs were obtained using an approximation from a Poisson distribution [20,22]. The Kaplan-Meier method was used to estimate actuarial probability of death amongst persons with a haematological cancer and competing risk analyses were done to estimate risk of dying from the index cancer, non-index cancer and noncancer causes.
Data on deaths were extracted from SEER 9 registries databases which continually code death trends from diverse causes by calendar year of cancer diagnosis from 1975 to 2016. In the trend analyses, we restricted follow-up interval to avoid bias from different intervals of follow-up time for persons diagnosed in different periods. For example, subjects diagnosed 1975-1979 were followed to 1984 (an additional 5-year minimum duration), those diagnosed 1980-1984 were followed to 1989, and those diagnosed in 2010-2014 were followed to 2016. All deaths during follow-up were recorded. Cancers diagnosed 2015-2016 were excluded because of inadequate follow-up. Joinpoint Software was used to fit the trends of deaths from index cancer, non-index cancer and noncancer causes [23]. Joinpoint Software generated one or more curves based on a log-linear regression model and so-called joinpoints, which  were used to connect adjacent curves to generate a continuous curve spanning the study interval. Percentage change per 5 years for each trend was estimated and integrated to generate an average 5-year percentage change for the entire trend. We describe risk of death from non-cancer causes as a function of haematological cancer-type and interval from haematological cancer diagnosis. Tests were two-sided and P-values < 0.05 were considered statistically significant. Analyses were done with SEER*Stat software version 8.3.6, the R version 3.63 statistical software, and Joinpoint Regression Software version 4.7.0.0 [15,23,24].

Causes of death by calendar interval
We analyzed death trends by assessing deaths from index cancer, other cancer or non-cancer causes by cancer type (Fig. 1). In persons diagnosed in 1975-1979, index-cancer deaths accounted for 65 percent of deaths, 2 nd cancer for 15 percent and non-cancer deaths, 20 percent. In persons diagnosed in 2010-2014, indexcancer deaths decreased to 62 percent (p = 0.04), 2 nd cancers decreased to 12 percent (p = 0.04) and non-cancer deaths increased to 26 percent (p = 0.02).
Persons most likely to die from their index cancer were those with PCM, ALL and AML. The proportion of deaths from index cancer remained stable or changed slightly for ALL (p = 0.5). For AML, there was a relative increase in index-cancer deaths beginning in 1985 (1975-1979 to 1985-1989 Table S1).

Causes of death by cancer type
The all-cause SMR for non-cancer deaths was 2.51 (2.49, 2.52) compared with the general population. Compared with cause-specific mortality in the general population, risks were highest for infections and kidney diseases, with cause-specific SMRs of 6.51 (6.42, 6.61) and 3.46 (3.34, 3.58), respectively. There were also increased risks of death for pregnancy, childbirth and puerperium complications (cause-specific SMR = 17.7 [13.6, 23.2]) ( Table 2).

Cause of death per interval after diagnosis
The interval immediately after a haematological cancer diagnosis had the highest risk of index-cancer-specific deaths except for CLL (Fig. 5). Subsequently, the likelihood of deaths from non-cancer causes increased in long-term survivors. In CLL, non-cancer causes were the leading cause of death regardless of post-diagnosis interval (Fig. 5). Persons with cancer had higher SMRs compared with the general population throughout the follow-up interval (Fig. 6). Most cancer types had high all-cause SMRs of non-cancer deaths in the 1st year after the haematological cancer diagnosis. All-cause SMRs decreased slightly during the next 4 years and then either decreased slightly or increased thereafter. For some non-cancer causes, the cause-specific SMR was highest in the 1st year after diagnosis including infections and external injuries. For other non-cancer causes, the cause-specific SMR in the interval after 5 years after haematological cancer diagnosis was like that or less than in the 5 years following diagnosis including CVD and respiratory and kidney diseases. People with newly-diagnosed ALL and AML had the highest all-cause SMRs of non-cancer deaths in the 1 st year after diagnosis. We also observed extremely high cause-specific SMR for infections and external injuries in persons with newly-diagnosed HL whereas an extremely high causespecific SMR for respiratory disease was seen in long-term HL survivors. Trends are displayed in Fig. 6.

DISCUSSION
We analyzed causes of death amongst 686,867 persons with haematologic cancers registered in SEER 1975-2016. Non-cancer deaths now exceed index-cancer deaths in CLL and will likely surpass index-cancer deaths soon in HL and CML. Causes of death varied by era of cancer diagnosis, cancer type, interval after diagnosis to death and other co-variates, including age and race.
As cancer therapy has improved, new cancers and non-cancer causes of death are increasing in long-term cancer survivors. We found risk of death from the index cancer, another cancer and non-cancer causes has changed for some but not all haematological cancers. As people with haematological cancer survive longer, risks of death from a new cancer and those from age-related non-cancer causes increase. There are many reasons why death from an index haematological cancer might decrease, including earlier diagnosis, better therapies, increased access to health care and others including changes in how deaths are coded and reported. Our analyses focused on characterizing cause of death amongst people with haematological cancer, not elucidating the reasons of changes which are likely complex. Such an analysis requires detailed subject-level data which are generally unavailable in population-based cancer registries such as SEER.
Our study is comprehensive but has several limitations. First, it is descriptive and retrospective. Second, cause of death may have been mis-classified because of inaccurate coding of death certificates in different intervals. This is an unlikely cause of substantial bias [25][26][27][28]. Rules for coding deaths changed during our study interval. It is also possible there might be a difference in accuracy of the cause of death in different cancers and non-cancer causes. Nevertheless, the SEER programme used standardized data collection procedures to ensure accuracy [28]. This approach has been validated [25-27, 29, 30].

CONCLUSION
Risk of death from an index cancer, new cancer or non-cancer causes varied by haematological cancer type, era of diagnosis, interval after diagnosis and other co-variates including age and race. Non-cancer deaths now exceed index-cancer deaths in CLL and will become dominants soon in HL and CML. Index-cancer deaths remained dominant cause of death amongst persons with PCM, ALL and AML in the 2010s. Changes in causes of death amongst people with different haematological cancers suggest distinct strategies to improve survival, such as increased screening for new cancers or non-cancer comorbidities. Non-cancer deaths might be decreased, prevented or effectively treated, especially in persons with CLL, HL and CML. In other haematological cancers including PCM, ALL and AML, preventing death from the index cancer remains the most pressing goal.

DATA AVAILABILITY
Data can be accessed from the Surveillance, Epidemiology, and End Results (SEER) database at https://seer.cancer.gov/data/.