Evaluation of seasonality in the diagnosis of acute myeloid leukaemia among adults in the United States, 1992-2008

Abstract

Recent studies have suggested seasonal variation in the diagnosis of acute myeloid leukaemia (AML), and the aetiological role seasonal factors may play in this group of haematological neoplasms remains unclear. We evaluated potential seasonality of AML diagnosis among adults. Cases included were ascertained from the Surveillance, Epidemiology, and End Results (SEER) 13 registries from 1992-2008. Chi-square analysis for heterogeneity and multiple Poisson regression using parametric harmonic modelling and bootstrap testing were used to detect possible monthly variation. Months of peak diagnoses were December and January, although some variation was present by sex and age. Heterogeneity across months was statistically significant (P < 0.001). In stratified analyses, monthly variation was detected only among males (P = 0.009) and in cases aged 65 years and older (P = 0.031). Poisson regression found no seasonal effect among all cases when fit to the sinusoidal model (P = 0.110). However, similar variation among males (P = 0.009) and cases aged 65 years and older (P = 0.018) was present. There is growing evidence of seasonality in AML diagnosis, particularly among older persons and men. Investigation of specific seasonal risk factors would be informative in explaining the aetiology behind the observed variation.

INTRODUCTION

Epidemiological inquiry into the periodicity of leukaemia onset has a long history and the aetiological role that seasonal factors may play in this group of haematological neoplasms remains unclear. The presence of seasonality in the diagnosis of acute myeloid leukaemia (AML) could possibly be an illustration of a seasonally incited aetiological event leading to onset and presentation of leukaemic disease, such as environmental factors or an infectious agent. Several studies have suggested a seasonal variation in the presentation and diagnosis of leukaemias and lymphomas (Douglas, et al 1999, Eatough 2002). These studies suggested a single peak incidence in early spring months (February/March) and a minimum of case diagnoses in late summer (August/September). However, the literature examining this possible seasonal effect is not entirely consistent, has not been updated in recent time periods, and mostly examines childhood cancers or rare subtypes in adults (Badrinath, et al 1997, Douglas, et al 1999, Eatough 2002, Harris, et al 1987, Higgins, et al 2001, Ross, et al 1999, Timonen 1999, Walker and van Noord 1982, Westerbeek, et al 1998).

The plausibility of a seasonal biological mechanism for leukaemia has arisen from research on temporal changes in immunological factors in humans. Seasonal variations in mitogenic responses and in the quantity of circulating lymphocytes, neutrophils, CD4 and CD8 cells, and interleukin 6 have been reported (Boctor, et al 1989, Maes, et al 1994, Nelson and Drazen 1999). Specific seasonal variations include a peak in lymphocyte aryl hydrocarbon hydroxylase activity in summer months (Paigen, et al 1981) and increased number of circulating B cells in winter months (Maes, et al 1994). Such data suggest, although stand to be further explained by additional research, possible biological factors underlying seasonal variation in haematological malignancy diagnoses. Given the availability of data from more recent time periods with greater diagnostic specificity and expanding statistical approaches to detecting seasonal trend, seasonality in the presentation of leukaemia subtypes warrants further study and would provide a basis for hypotheses of potentially causal risk factors for leukaemias.

The purpose of our study was to investigate potential seasonal trends in AML diagnoses among adults using data from the Surveillance, Epidemiology, and End Results (,) 13 registries between the years 1992 to 2008. We hypothesized that AML incidence observed over the study period would have a single annual peak in the winter months. Our study adds to existing literature by investigating seasonality in a geographically diverse population-based sample of adult AML cases and inclusion of a recent time period while employing novel statistical approaches recently recommended to test serial count data in SEER (Hunsberger, et al 2002). Using the large amount of data available in SEER will provide evidence building on previous studies in order to confirm or refute possible seasonal variation.

MATERIALS AND METHODS

Study population

We included cases of AML classified according to the World Health Organization (Vardiman, et al 2002) including acute monocytic leukaemia in adults aged 25 years and older, diagnosed from 1992 through 2008 and reported by 13 SEER registries. The SEER 13 registries cover central registries in five states – Connecticut, Hawaii, New Mexico, Iowa, and Utah – and the eight respective US reporting areas and tumour registries of Atlanta, Detroit, San Francisco-Oakland, Seattle-Puget Sound, Los Angeles, San Jose-Monterey, Rural Georgia, and Alaska Native territories (SEER 2011 [http://www.seer.cancer.gov]). We excluded cases without recorded month and year of diagnosis. Our study included 21,570 incident AML cases occurring over 422,933,387 person years. SEER collects patient-specific information from which we derived case characteristics, including age group (aged 25-44, 45-64, ≥65 years), gender (male, female), race (Non-Hispanic White, Hispanic, Asian or Pacific Islander, Black, American Indian or Alaska Native), and leukaemic subtype. Data collected from SEER has a high level of overall completeness and ascertainment in the respective reporting areas (Kim, et al 2011, Thoburn, et al 2007).

Statistical analysis

Age-standardized (2000 US standard population) incidence rates of AML were calculated using SEER*Stat statistical software (seer.cancer.gov/seerstat) version 7.1.0. Cases per month were standardized to months of equal length.

We used a chi-square test of homogeneity to test the hypothesis of no difference in number of diagnoses by month. Six separate tests were performed using data for all adult cases, all male and all female cases, and all adult cases in the age ranges 25-44, 45-64 and ≥65 years. The test shows the presence of any departure from a uniform distribution throughout the year. Next, Poisson regression models were used to determine the pattern of seasonality in AML incidence. This method fits sinusoidal (harmonic) models to the data, using observed counts as the outcome and expected counts and month of diagnosis as covariates. This approach, proposed by Hunsberger, et al (2002) is intended to test seasonality using parametric models for seasonal trends in serial count data, assessing possibly complex seasonal patterns by sequentially adding harmonic (sine and cosine) terms to the model. The model uses a penalized likelihood to choose the number of harmonics, and then tests for seasonality using a parametric bootstrap test. Incorporating the selection of harmonics in the bootstrap test thus controls the type I error rate. This method models the number of diagnoses in month i, Y_i, as Poisson with mean μ_i, using the following model for the conditional mean, μ_i^c:

$$\left\{Y_i\right\}{{}^n}_{i=1}=\text{log}{{\mu }_i}^{\mathrm{c}}={\beta }_0+{\beta }_1{t}_i+{\Sigma }_{k=1,p}\{{\beta }_{ck}\text{cos}\left(2k\pi m_i\right)+{\beta }_{sk}\text{sin}\left(2k\pi m_i\right)\}+⍰\left[\text{log}\left({\mathrm{y}}_{i\text{-}1}\right)\text{-}{\beta }_0\text{-}{\beta }_1{t}_{i\text{-}1}\text{-}{\Sigma }_{k=1,p}\{{\beta }_{ck}\text{cos}\left(2k\pi m_{i\text{-}1}\right)+{\beta }_{sk}\text{sin}\left(2k\pi m_{i\text{-}1}\right)\}\right](\text{Hunsberger},etal2002)$$

Here, t_i is a variable indicating the sequence of times at which counts were taken; m_i indicates the calendar month and takes positive numeric values between 0 and 1 with 1/12 for January, 2/12 for February, and so on, ending with 12/12 for December, respectively. The number of harmonic components, p, is selected from the data. The parameters, β₀, β₁, β_ck, β_sk and p, are estimated from the data. β₀ is the intercept, β₁ is the coefficient of a linear time trend and β_ck and β_sk are the coefficients of the harmonic model of order p. Theta, ⍰, is an autocorrelation parameter that accounts for dependence on the previous month's count. Seasonality is tested by jointly testing β_c = β_s = 0. Such parameterization forces the seasonal component to follow a sinusoidal curve, adding in higher-order harmonic terms for multiple annual peaks and troughs for modelling flexibility. The order of the harmonic model was not allowed to exceed six by restricting the highest order to p = 5. If p ≥ 6, the model becomes completely saturated with monthly harmonic terms (each month ≥ 1 peak or trough). Although Poisson harmonic regression model is for the conditional mean, the marginal mean, μ_i, is approximately equal to μ_i^c so that inferences based on the model are approximately marginal.

The month when the maximum incidence occurred was estimated, applying models of a single annual peak within the period of a year, and was adjusted for SEER reporting registry. The significance of the sinusoidal model assumption was evaluated with post-estimation, Pearson chi-square tests for goodness-of-fit and AIC/BIC (Akaike information criterion/ Bayesian information criterion) criteria. Stratified analyses were also performed for gender and age groups. Statistical significance was taken to be P < 0.05. All analyses for heterogeneity and seasonality of AML diagnoses were performed using Stata version 12 (Stata Statistical Software: Release 12. College Station, TX: StataCorp: LP).

RESULTS

Descriptive characteristics of the AML cases are reported in Table I. More cases were aged 65 years and older (59.2%) than aged 25 to 44 years (13.3%) or 45 to 64 years (27.4%). The majority of cases were non-Hispanic White (66.9%); smaller proportions of cases were Hispanic (12.0%), Asian or Pacific Islanders (10.6%), and Black (9.7%). A slightly greater proportion of cases were male (53.7%) than female. SEER registries in metropolitan areas and those of high average estimated population composed the greatest proportion of case reporting, such as Los Angeles (22.2%) and Detroit (12.4%). Conversely, areas of lower population estimates reported the least number of cases among the registries, including Alaska Natives and Rural Georgia. Of the 21,570 AML cases identified between 1992 and 2008, a minority were acute monocytic leukaemia (7.3%) (data not shown).

Table I.

Incident adult AML diagnoses by age, gender, and race/ethnicity, 1992-2008, N = 21,570

Age	N	(%)
25 to 44 years	2,878	13.34%
45 to 64 years	5,918	27.44%
65+ years	12,774	59.22%

Gender	N	(%)
Male	11,592	53.74%
Female	9,978	46.26%

Race / Ethnicity	N	(%)
Non-Hispanic White	14,436	66.86%
Hispanic	2,591	12.00%
Asian or Pacific Islander	2,287	10.59%
Black	2,092	9.69%
American Indian or Alaska Native	164	0.76%

Registry	N	(%)
San Francisco - Oakland SMSA	2,245	10.41%
Connecticut	2,114	9.80%
Detroit (Metropolitan)	2,675	12.40%
Hawaii	724	3.36%
Iowa	2,254	10.45%
New Mexico	872	4.04%
Seattle (Puget Sound)	2,467	11.44%
Utah	927	4.30%
Atlanta (Metropolitan)	1,193	5.53%
San Jose - Monterey	1,208	5.60%
Los Angeles	4,791	22.21%
Alaska Natives	27	0.13%
Rural Georgia	73	0.34%

Year	N	(%)	Year	N	(%)
1992	1,037	4.81%	2001	1,418	6.57%
1993	1,080	5.01%	2002	1,315	6.10%
1994	1,093	5.07%	2003	1,357	6.29%
1995	1,200	5.56%	2004	1,270	5.89%
1996	1,112	5.16%	2005	1,304	6.05%
1997	1,228	5.69%	2006	1,325	6.14%
1998	1,314	6.09%	2007	1,392	6.45%
1999	1,284	5.95%	2008	1,451	6.73%
2000	1,390	6.44%

AML, Acute myeloid leukaemia; SMSA, Standard metropolitan statistical area.

The observed monthly incidence rates of AML diagnoses over the study time period were based on a standardized US Census 2000 population of 422,933,387 person-years, and monthly incidence rates ranged from 4.2 to 4.8 per million person-years (Table II). These observed rates were similar to an expected overall incidence rate of 5.0 per million person-years. The months of AML case diagnosis in which annual peaks in the overall count were observed were December and January, although in stratified analyses of gender and age other potential peaks emerged (see Table II). Testing the hypothesis of no difference (uniform distribution) of monthly AML diagnoses among all cases in this sample with a chi-square test was statistically significant (P < 0.001). In stratified analyses, a non-uniform distribution or possible seasonal effect among males was detected (P = 0.009) whereas no such effect was found among females. Among age groups, a departure from a uniform distribution of cases was detected only in the cases aged 65 years and older (P = 0.031), again, with peaks in December and January. By comparison, pronounced peaks in December and January were not observed among females and younger age groups (Table II).

Table II.

Cases of adult AML per calendar month, standardized incidence rates, and heterogeneity tests, 1992-2008

All adult AML cases and age-adjusted rates per million												Heterogeneity tests
May	June	July	Aug	Sept	Oct	Nov	Dec	Jan	Feb	Mar	April	X 2	P-value⍰
Cases												31.29	P < 0.001
1,774	1,841	1,788	1,822	1,724	1,812	1,761	1,844	1,878	1,651	1,862	1,813
Rate
4.53	4.69	4.56	4.65	4.40	4.63	4.50	4.72	4.80	4.21	4.74	4.63

All male, adult AML cases and age-adjusted rates per million												Heterogeneity tests
May	June	July	Aug	Sept	Oct	Nov	Dec	Jan	Feb	Mar	April	X 2	P-value⍰
Cases												25.18	P = 0.009
943	959	952	964	909	973	955	986	1,057	893	1,011	991
Rate
5.57	5.62	5.70	5.72	5.41	5.87	5.70	5.94	6.35	5.32	6.00	5.89

All female, adult AML cases and age-adjusted rates per million												Heterogeneity tests
May	June	July	Aug	Sept	Oct	Nov	Dec	Jan	Feb	Mar	April	X 2	P-value⍰
Cases												16.38	P = 0.127
831	882	836	858	815	839	806	858	821	758	854	822
Rate
3.77	4.01	3.82	3.91	3.71	3.84	3.65	3.91	3.74	3.45	3.86	3.73

All adult AML cases and age-adjusted rates per million, aged 25-44 years												Heterogeneity tests
May	June	July	Aug	Sept	Oct	Nov	Dec	Jan	Feb	Mar	April	X 2	P-value⍰
Cases
232	244	240	257	210	231	219	257	256	227	261	244
Rate												12.98	P = 0.295
1.13	1.18	1.18	1.26	1.03	1.13	1.07	1.26	1.25	1.11	1.27	1.19

All adult AML cases and age-adjusted rates per million, aged 45-64 years												Heterogeneity tests
May	June	July	Aug	Sept	Oct	Nov	Dec	Jan	Feb	Mar	April	X 2	P-value⍰
Cases
481	525	506	498	512	494	479	485	490	442	520	486
Rate												12.57	P = 0.322
3.40	3.71	3.56	3.52	3.62	3.50	3.37	3.44	3.46	3.12	3.67	3.42

All adult AML cases and age-adjusted rates per million, aged 65+ years												Heterogeneity tests
May	June	July	Aug	Sept	Oct	Nov	Dec	Jan	Feb	Mar	April	X 2	P-value⍰
Cases
1,061	1,072	1,042	1,067	1,002	1,087	1,063	1,102	1,132	982	1,081	1,083
Rate												21.19	P = 0.031
14.53	14.70	14.29	14.63	13.74	14.88	14.57	15.11	15.53	13.47	14.83	14.84

^⍰X² test for homogeneity to test the hypothesis of no difference (adjusted, uniform distribution) in number of diagnoses by month AML, Acute myeloid leukaemia

The fitted trend line of the harmonic regression model for all cases (Figure 1) suggested some agreement with the peak period months based on monthly incidence rates (Table II), with higher incidence in December and January; however, another peak was suggested in June and July. Although heterogeneity for all cases was highly statistically significant using the chi-square test (Table II), the monthly heterogeneity was not significant when fit to the sinusoidal model (P = 0.110). When considering modelling of subgroups, interaction terms were added to the overall model of all cases. These were found to be statistically significant for heterogeneity between age groups and sexes. Thus, stratification of monthly rate data to the Poisson harmonic regression model by sex (Figures 2a and 2b) described significant seasonal variation among males (P = 0.009) and no such variation among females (P = 0.126). Of the age groups, the sinusoidal model fit well among adult AML cases aged 65 years and older (P = 0.018; Figure 3); no significant variation was found within other age strata of ages 25-44 years (P = 0.254) nor ages 45-64 years (P = 0.381) (data not shown). These Poisson harmonic regression results largely agreed with the results comparing heterogeneity in monthly rates by the chi-square tests of goodness-of-fit (Table II). In particular, both methods found that seasonal variation was most pronounced in males (Table II and Figure 2a) and in those aged 65 years and older (Table II and Figure 3), with peaks in December and January.

Figure 1.

SEER 13 Diagnoses of adult AML with Poisson harmonic regression trend line. Poisson harmonic regression, sinusoidal model post-estimation P-value (Pearson X² goodness-of-fit)

Figure 2.

SEER 13 Diagnoses of AML by gender with Poisson harmonic regression trend lines. Poisson harmonic regression, sinusoidal model post-estimation P-value (Pearson X² goodness-of-fit)

Figure 3.

SEER 13 Diagnoses of AML in ages ≥ 65 years with Poisson harmonic regression trend lines Poisson harmonic regression, sinusoidal model post-estimation P-value (Pearson X² goodness-of-fit)

Given the variability in trend lines and strata of females and age groups 25 - 44 years and 45 - 64 years, our assumption of a single annual peak in the harmonic Poisson model was examined further. Higher-order models allowing the possibility of multiple peaks, such as a bimodal trend, were tested over all cases and within strata of age and gender. All models were evaluated for significant trends and using post-estimation statistics. From these analyses, goodness-of-fit of the models were not improved compared to those examining the significance of a possible single annual peak.

DISCUSSION

Analyses of serial counts data of adult AML monthly diagnoses for years 1992 to 2008 from the SEER 13 registries provide evidence that suggests presence of some seasonal effect, with a peak during the months of December-January. Such seasonal effects were most consistent and pronounced among males and persons aged 65 years and older. This study contributes to the existing literature by providing a more broad examination of adult AML cases from a wider range of population-based registries in the US in a more recent time period than previous reports on this subject. Based on other studies of SEER incidence rates, we applied a Poisson harmonic model and bootstrap tests of seasonal trend in a large amount of SEER data considered to have similar or greater power as alpha-level procedures versus other approaches used to investigate possible seasonal trends (Hunsberger, et al 2002).

While there are no directly comparable studies to our analysis, results from other studies using cancer registry data from the US and Europe have reported on cases ascertained in varying periods between the 1940s and 1990s. An epidemiological study (Douglas, et al 1999) in the UK examined the seasonality in the diagnosis of childhood and adult non-Hodgkin lymphoma (NHL) and leukaemia cases in the years 1984 to 1993, including AML, using methods similar to those presented here. Analyses of AML diagnoses in this study were stratified by gender but not age. Cosinor analysis, which did not account for correlation for cases aged 0 to 79 years, did not detect heterogeneity in the distribution by month in males (P = 0.84) or females (P = 0.29), or in all cases analysis (P = 0.87). Data from this study, however, were suggestive of a seasonal trend when performing normal approximation to the Poisson distribution. In stratified analyses, peak diagnostic months were identified to peak in February and January for males and females, respectively. Detectable seasonal trends for childhood acute lymphoblastic leukaemia (ALL) and NHL were observed, while no definitive evidence of seasonal variation in AML diagnosis was present.

More recently, an analysis (Eatough 2002) in the UK examined seasonality in acute monocytic leukaemia, a rare AML subtype, using data combined from four cancer registries. Cases of acute monocytic leukaemia in the UK were ascertained from the UK Office of National Statistics for England and Wales, 1974 to 1998, and in Britain for the years 1946 to 1960 among cases aged less than 45 years. In the US, 306 childhood and adult cases were ascertained from the Third National Cancer Survey, SEER, and the Central Cancer Patient Data System from the years 1969 to 1971, 1973 to 1977, and 1971 to 1981, respectively (Eatough 2002). By chi-square analysis, the combined data including 3000 cases provided strong evidence of a non-uniform distribution of cases diagnoses by month (P < 0.005). The peak months of February and March and trough months of August and September for monocytic leukaemia diagnosis observed by Eatough (2002) differed from peak and trough months found for all AML cases in our study. In our sample, cases of acute monocytic leukaemia composed a small proportion of AML cases. When stratified by subtype, no significant variation from a uniform distribution was observed in the monocytic cases alone (N=1575 cases, P = 0.432 by chi-square analysis) (data not shown).

As hypothesized, evidence of seasonal variation was present in our analysis, particularly among subgroups known to be at higher risk of AML. The significance of the flexible parametric harmonic Poisson model of an annual peak period in our results provides evidence supporting a seasonally mediated exposure inciting AML onset, such as infection or allergy. Chronic viral infection is implicated as a causal factor in many cancers. Carriage of a viral leukaemic pathogen, correlating with host immunological changes and diagnosis of incident leukaemias, has been suggested (de Martel and Franceschi 2009, Greaves 1997, Pagano, et al 2004) One reasoning behind such seasonal variation in leukaemia diagnosis is that peak incidence occurs when pathogens are present in a susceptible population and other important immunological factors act in concert. Subsequently, acute infection or reactivated chronic infection could be responsible for induction of leukaemia, symptoms, and incident diagnosis (Dowell 2001).

Separately, an alternative explanation to be considered is that a peak is observed during these winter months as a result of increased surveillance (Sackett 1979). During these periods, people are more likely to be seen by providers for seasonal conditions such as influenza. Such complaints may lead to tests or procedures leading to the diagnosis of AML (Meyers et al 2005). Perhaps supporting such an explanation is the sharp decline that is observed in February following the December-January peak. Such an observation would be present if cases that would have otherwise been diagnosed in February were being detected in the two months prior because of more frequent provider visits, for example (Haut and Pronovost 2011). Other investigators have suggested that observed seasonal variation in cancer might be the result of lower diagnosis rates during vacation months, such as summer (Lambe, et al 2003). We examined the possibility of surveillance bias by drawing comparisons to studies examining seasonality in other adult haematological malignancies. Seasonality was not detected for adult Hodgkin lymphoma (Douglas, et al 1998) or an adult ALL subgroup (Douglas, et al 1999) in two epidemiological studies, although their methods differed from our approach. Further, observed peak months of diagnosis for these cases were inconsistent with those for other cancers in the respective populations. If surveillance bias were strongly present, we would expect similar overlap of peak or trough months for various cancers. We additionally evaluated the possibility of surveillance bias in our data by a post-hoc comparison of our results for adult AML with an analysis of seasonality in another type of leukaemia commonly occurring in adults – chronic lymphocytic leukaemia (CLL). Similar to AML, CLL is most common in men and in the elderly (Hernandez, et al 1995, Smith, et al 2011). If surveillance bias accounted for the seasonal pattern we observed for AML, we would expect to see a similar pattern for CLL (although even in this situation a true aetiological seasonal pattern could not be ruled out for both types of leukaemia). When analysing a sample of CLL cases (N = 25,942) in the SEER 13 registries from the same age range and years as our AML analysis, we found strong evidence of a monthly variation in diagnosis. The peak month of diagnosis over the study period for all cases, among both males and females, and all age groups was June (data not shown), which differed from the peak months of December-January observed for AML. These results provide evidence against surveillance bias as the sole explanation for the seasonal variation we observed in AML diagnoses.

It should be acknowledged that important limitations are relevant to this study. First, although reporting in the SEER 13 registries is considered to be of high quality, undergoing validity testing and evaluation bi-annually, only the month of diagnosis (i.e. histological confirmation) was collected in the ascertainment of our cases. While our use of the publicly available SEER data allowed inclusion of a large, geographically defined sample, it limited our ability to investigate person-level data that may confound seasonal trends in diagnoses. Specifically, we were unable to examine data pertaining to incidence of first symptoms, frequency of healthcare utilization and access, or the presence of comorbid conditions. All of these covariates could affect the duration between disease onset and diagnosis. It is also important to note that smaller samples of cases in subgroups of age (younger) and gender (female) could contribute to Type II error, preventing us from observing a seasonal effect in those subgroups. If lack of power entirely explained no variation in these subgroups, we would probably expect a similar pattern in peak months to that of males and older persons and this was not observed in these data. Given the limitations, our study of seasonal patterns of AML diagnoses across the SEER registries is intended to be hypothesis generating rather than providing evidence for causal inference. Clearly, individual-level studies are needed to determine if the growing amount of evidence for seasonality is the result of seasonal variation in aetiological factors or is rather due to diagnostic bias or other uncontrolled factors.

In conclusion, this analysis and other recent articles described above that encompass the experience of several decades of AML diagnosis surveillance indicate an increasing capacity to evaluate evidence of seasonal trends. Evidence from this analysis demonstrates some agreement between past studies suggesting seasonal variation across different registries and over a wide range of years of observation. Stratification of groups by important covariates (age, gender, comorbidity, environmental exposure) and use of time series methods that are able to control for type I error will probably be important tools in future studies when evaluating seasonal variation in large sets of serial count data in SEER and other cancer registries internationally. Investigation of specific seasonal risk factors would be informative in future studies to explain the aetiology behind the observed variation.