Computed tomography scan radiation and brain cancer incidence

Cohort and Data Linkage

As previously reported, deidentified electronic Medicare records of all Australians aged 0–19 years between 1985 and 2005 were linked to national cancer incidence and death records held by the Australian Institute of Health and Welfare, with a follow-up to 2007.¹³ The present report is based on the extended follow-up to December 31, 2012, and on new organ dose estimates based on retrospective organ dosimetry for each CT exposure recorded between 1985 and 2005.^12,13 Patients entered the cohort on the earliest date that they could have been exposed to a Medicare-funded CT scan, (the latest of January 1, 1985, date of birth, or date first known to Medicare). Exit date was the earliest of either December 31, 2012, their date of death, or the date of a first cancer diagnosis. An index of socioeconomic disadvantage (SEIFA) was included as a potential confounder; approximately one million persons with a missing SEIFA were excluded.¹⁴

Compliance with Ethical Standards

The study was conducted in accordance with National Health and Medical Research Council Guidelines on Human Experimentation and with relevant government legislation. Access to linked data, without individual subject consent, was approved by data custodians from all state and territory jurisdictions in Australia, and by human ethics committees from the University of Melbourne, the Australian Institute of Health and Welfare and jurisdictions. All subjects were deidentified and birth dates were rounded to the nearest month to protect confidentiality. Access to unit records was restricted to further reduce the possibility of reidentification.

Dosimetry

Organ doses were estimated using the NCICT program developed at the National Cancer Institute, National Institutes of Health in the United States.^12,15 To provide the necessary inputs, we used Medicare billing codes assigned for over 200 different CT scan types, together with CT technical parameters derived from Australian national surveys, clinical protocols, regulator databases, and literature to reconstruct volumetric CT dose index (CTDI_vol) estimates for each imaging procedure, by year of service and patient age.¹² Multiple contrast phases for CT scans were taken into account. Brain doses were estimated for all recorded CTs (head and neck, spine, chest, abdomen, and extremities).

Exposure Lagging

Dates of CT exposures were lagged to minimize reverse causation bias and to account for latency. The lag period (L) was the length of time following exposure before the exposed individual was transferred from the unexposed to the exposed cohort on the “transfer date” (see Figure 1). Only persons whose first lagged CT scan date was prior to their exit date contributed follow-up years to the exposed cohort. Years before the transfer date (including lag years) were counted as unexposed person-years.

The cumulative brain dose at time t was the actual cumulative dose at time t−L, where L is the lag period.^16,17 We routinely used a lag period of 2 years, but also used 5-year lag periods for gliomas to allow for slower growth rates and clinical management strategies; (see Supplementary Material Section 1.1 and 1.2).

Exclusions

Prior to the application of exclusions, the dataset was composed of 11 802 846 persons. Persons with a missing SEIFA score were excluded due to a likelihood of a missed linkage (n = 996 590). Persons with a first nuclear medicine that was greater than 2 years before the diagnosis or radiotherapy exposures greater than 100 days prior to diagnosis were excluded (n = 273 979 and n = 26 520 respectively, noting that there is overlap). Persons with more than 7 CT scans (n = 452) were also excluded because of the possibility that larger numbers of scans could have been ordered after diagnosis for the management of the few cancers that would have been missed in the linkage process. Thus, after excluding 1 278 004 persons there was a total of 10 524 842 people.

Outcomes

The primary outcome was the diagnosis of intracranial brain cancer, as reported by the Australian Cancer Database.^18–20 Diagnoses were coded in ICD 10, with brain cancer subtypes identified using the World Health Organization ICD-O-3 classification (see Supplementary Material Section 1.3 and Supplementary Table 1).²¹ Brain cancers were broadly categorized as gliomas, medulloblastomas (MB)/primitive neuroectodermal tumors (PNETs). Gliomas are comprised of high-grade gliomas and low-grade gliomas. high-grade gliomas included World Health Organization grade III and IV gliomas. low-grade gliomas included World Health Organization grade I and II gliomas. A small group of miscellaneous cancers occurring in the cranium such as neuroblastomas and chordomas was included in the overall brain cancer counts as exposure to ionizing radiation may have increased tumor risk. Meningiomas, as benign tumors, were not recorded. We censored nervous system cancers outside the cranium because the organ dose to the brain would not be relevant. The few cancers from non-nervous system tissue within the cranium (lymphomas, blood vessel tumors, sarcomas, teratomas, and benign neoplasms) were also censored.

Statistical Analyses

We used a person-year dataset stratified by age at exposure (as in the LSS²²), attained age, gender, socioeconomic categories (7 quantiles of the SEIFA index), year of birth (pre-1985, 1985–1995, and 1995–2005), and dose categories (based on the quartiles of the person-year weighted brain doses). In sensitivity analyses, we also tested 3 and 5 dose categories (based on percentiles) and different lag periods (0 to 20 years). As in our reanalysis of the Japanese atomic bomb survivors cohort,²³ the incidence rate ratio was estimated by stratified Poisson regression with person-year weighted brain dose as the predictor and brain cancer counts as the outcome. The excess relative risk (ERR) was estimated as an incidence rate ratio of −1. Confidence intervals for the attributable fractions in the exposed (AFE) and in the population (AFP), and the number needed to harm, were obtained using the substitution method.²⁴ Dose was time-varying to reflect increasing cumulative doses over time. We selected the model with the lowest Bayesian Information Criterion value. All analyses were performed on Stata/MP 16.0.

Descriptive Statistics

This study included 10 524 842 people, with 4472 incident brain cancers and a total of 235 643 748 follow-up years (for more details, see Supplementary Section 2 and Supplementary Table 2). Overall, less than 5.8% of the cohort were exposed to a CT scan in the years 1985–2005 at ages 0–19 years. The crude incidence rate for brain cancers was 1.86 per 100 000 person-years. The average follow-up was 22.4 years for the whole cohort, and 13.5 years for CT-exposed after the 2-year lag period had expired. With the 2-year lag period, 717 095 CT scans were included for 611 544 CT-exposed persons with 8 259 452 years of follow-up time; 472 391 persons (77%) were exposed to one or more brain scans. Females were 22% (95% CI 18%–27%) less likely than males to be diagnosed with brain cancer.

With a 2-year lag, 237 exposed persons were diagnosed with incident brain cancer. Table 1 describes the estimated brain doses in the exposed population. The median cumulative brain dose was 32 milligray (mGy) (Q1 1.4, Q3 63). The average cumulative brain dose delivered was 44 mGy (standard deviation 48 mGy). For 95% of those exposed, the total brain dose was less than 127 mGy. For the 536 335 brain CT scans, the median brain dose per scan was 38 mGy (IQR 59 mGy).

Exposed Versus Unexposed and Excess Risk Per CT Scan

The risk of developing brain cancer for persons exposed to CT scan radiation before the age of 20 years was 67% greater than the risk for unexposed persons, after adjustment for age, gender, year of birth, and socioeconomic index. Thus, the ERR for brain cancer was 0.67 (95% CI 0.40–1.98). For persons exposed to a single CT scan before the age of 20 years, the risk of brain cancer was 43% greater than for unexposed (see Table 2); this corresponds to an ERR of 0.43 (0.18–0.73). For those who received 3 or more scans (3–7 scans), the ERR was 3.7 (1.83–6.8). For single brain scan exposures, the ERR was 0.54 (0.26–0.88) and for 3 or more brain scans the ERR was 5.29 (2.44–10.5). As expected, the effect size was largest with a zero lag (Figure 2), 
and it declined progressively at longer lag periods.

Dose–Response

Brain cancer risk increased with cumulative radiation dose to the brain. For all brain cancers the excess risk increased by an average of 80% per 100 mGy of dose, corresponding to an ERR/100 mGy of 0.80 (95% CI 0.54–1.06; see Table 4) with a 2-year lag. Table 4 also summarizes dose–response relationships by histological subtypes. For example, for high-grade gliomas, the ERR/100 mGy was 0.73 (95% CI 0.20–1.27) with a 2-year lag. The estimated (linear) effect for the exposed cohort only (unexposed cohort excluded) was an ERR/100 mGy of 1.26 (95% CI 0.86–1.66). Adjustment for age at exposure did not improve the fit of any of our models.

Attributable Fraction in the Exposed (AFE), Excess Absolute Risk, Attributable Fraction in the Population (AFP), Number Needed to Harm,

Of brain cancers in our cohort that followed CT exposures in childhood by more than 2 years, 40% were attributable to CT scan radiation (ie AFE = 40% with 95% CI 28.8, 49.5%). Brain cancer incidence increased by 1.16 (95% CI 0.91, 1.41) per 100 000 person-years following CT scan exposure. This corresponded to an excess absolute risk of 2.03 (95% CI 1.78, 2.28) per 100 000 follow-up years at a dose of 100 mGy. Because of the relatively low rate of CT scan exposure in our cohort, less than 4% of all brain cancers in the cohort were attributable to CT scan radiation (ie, AFP = 3.7% [95% CI 2.3, 5.4%]). For an average follow-up of 13.5 years, 6391 (95% CI 5255, 8155) persons need to be exposed to cause one extra brain cancer (number needed to harm; see Table 3). 
At 100 mGy of exposure, 3643 persons need to be exposed (95% CI 3243, 4155) to cause one extra brain cancer.

For further details, with a breakdown by histological subtype, see Supplementary Material Section 2 and Supplementary Tables 4-6.

Sensitivity Analysis

If persons with more than ten, or more than twenty, CT scans were excluded (instead of more than 7), the estimated ERR/100mGy remained stable at 0.62 (95% CI 0.47, 0.78; P < .001) and at 0.62 (95% CI 0.47, 0.76; P < .001) respectively, demonstrating the robustness of these findings. If the persons with missing SIEFA were included, the ERR for exposure was 0.65 (0.37, 0.98; P ≤ .001) and the dose–response ERR/100 mGy was 0.87 (95% CI 0.52–1.22) indicating the stability of these findings despite these exclusions.

For comparison purposes, we estimated the dose–response using Cox proportional hazards regression models with individual rather than grouped data. With sex, year of birth, and SEIFA included as covariates and using a 2-year lag period, the hazard ratio was 0.60 (95% CI 0.50, 0.69; P < .001) per 100mGy of radiation for all brain tumors; this estimate is very similar to the dose–response estimate using grouped stratified data (ERR/100mGy 0.62).

Strengths and Limitations

Our large cohort captured all CT scans funded on a fee-for-service basis by the universal Australian Medicare system for persons aged 0–19 years in the period 1985–2005. The new estimates presented here are based on individual dose estimates for each CT scan using the NCICT program, 5 more years of follow-up, and a 2-year lag period. These changes help to explain why our current risk estimates are less than our earlier provisional estimates from the same cohort using a 1-year lag period.⁷

Between 1985 and 2005 there were changes in practice and technology leading to reduced radiation doses. The organ dose calculations performed using the NCICT program to estimate doses did take into account the year of the scan, and thus scans occurring in the later years had lower organ dose values.¹²

The aggregate rate of CT exposure in our cohort is likely underestimated. Some children requiring specialist treatment attended clinics in state-funded hospitals where some CT scans were not billed to Medicare, leading to missed exposures. Furthermore, because Medicare records were only available from 1985–2005, our records could not include all exposures between ages 0–20 years for all cohort members.

In this work, we have not included estimates of radiation doses from conventional x-rays and fluoroscopic studies, and to minimize misclassification bias we intentionally excluded patients receiving radiation therapy and nuclear medicine procedures. Often, persons having a CT scan also had conventional x-rays, such as children with ventriculoperitoneal shunts had CT scans and x-rays to check the position of the catheter (postoperative review or if there was suspected dysfunction). Underestimation of brain doses is likely to be small, as plain x-rays rarely deliver substantial doses to the brain. For example, the median brain dose per CT scan was 38 mGy, and a chest x-ray would deliver much less than 0.1 mGy to the brain. Brain dose estimates were also subject to other sources of error. For example, for a short period, some brain scans were coded as being “with or without” contrast, so that it was not possible to decide whether contrast had actually been used for specific scans, which would double the dose for the few individuals concerned. We dealt with this problem by estimating the probability of contrast use from adjacent time periods when the contrast code was not ambiguous.¹²

The crude incidence rate of 1.86 per 100 000 person-years was lower than that noted in the Surveillance, Epidemiology, and End-Results database (SEER) which is 8.5 per 100 000 person-years.³³ This difference is likely due to the younger age of our cohort as the oldest person in our cohort was 47 years old in 2012. The median age at diagnosis for persons with a GBM in the SEER database was 64 years with an interquartile range (IQR) of 21 years.³⁴ We note that the exclusion of 1 278 004 persons may be a cause for concern. We demonstrated through a sensitivity analysis that the effects of removing these observations were minimal.

We found that the ERR tended to decline at longer lags (see Figure 2), and that with a lag of 2 years, the ERR declined progressively with time since exposure (results not shown). There is a tendency to attribute such time trends to confounding by indication, even though the time-related decline in ERR is also seen in the LSS, where it cannot be caused by such confounding. The decline of ERR with time since exposure is perhaps best explained by the direct effects of dose in depleting the pool of (genetically) susceptible individuals over time. Thus, causal pathways for effect modification or confounding by indication may be more complex than previously thought.

There is an ongoing debate regarding the effects of confounding by indication on dose–response effects. For confounding by indication to be present, persons with a predisposing factor would be more likely to receive a CT scan, and more likely to be diagnosed with cancer. In a preliminary study estimates of ERR associated with CT exposures were reduced by 17%–56% (based on a 2-year exclusion period), depending on the site of cancer, after adjusting the risk estimates for predisposing factors.⁹ A follow-up study found that CNS tumor risks appeared to increase with CT scan doses in patients without a predisposing factor, while they tended to decrease in children with predisposing factors, an argument against confounding by indication.³⁵ A recent study using the Aust-PERC cohort found minimal change to the point estimates before and after propensity score weighting to adjust for the likelihood of receiving a CT scan (control for confounding by indication).³⁶

The difference in radiation-related risks observed according to the presence or absence of predisposing factors might be explained by much higher mortality risks in persons with predisposing factors.³⁷ A high early non-cancer mortality caused by predisposing factors would decrease the number of children at risk of cancer in this group. A separate study using linked data from the NHS in Great Britain concluded that predisposing conditions do not bias the relationship between ionizing radiation and brain cancer risk and that persons with pre-disposing factors did not undergo more CT scans.³⁸ Thus, while it remains possible that persons with predisposing factors may bias the relationship, we believe that the effect of such a bias would be small due to the small proportion of persons with predisposing factors and the small amount of CT scans done on this population.^3,38 Overall, it is likely that the increase in brain cancer risk following CT scans in childhood is causal, and not due to reverse causation or confounding by indication.

Clinical Implications

Indeed, the future population risk may be greater in populations such as the United States, where CT scanning rates in children are now higher than they were in our cohort between 1985 and 2005.² Future cancer risks can be reduced by using low-dose CT or rapid MRI protocols for children who need multiple scans during childhood.^39–42 For example, the risk of brain cancer for those with ventriculoperitoneal shunts in childhood would be high as they receive an average of 3.3 CT scans in addition to several skull x-rays (brain dose range 0.1–2 mGy) and chest x-rays (brain dose <0.001 mGy).⁴³ Thus, practice changes are warranted, especially for children with ventriculoperitoneal shunts.