More than one million U.S. radiation workers and veterans have been identified through employment records, military records, and regulatory and commercial dosimetry databases. Examples of radiation-exposed cohorts evaluated to date include workers at Los Alamos National Laboratory (14), at the Tennessee Eastman Corporation Y-12 facility (18), U.S. military participants at above-ground nuclear weapons tests (13), nuclear utility workers (15), industrial radiographers (2,19), and medical radiation workers (17). The overall study approach and methods for mortality tracing, analyses, and dose reconstruction have been described (14,34–37). High- and low-LET radiation dose (e.g., from photons, radionuclide intakes, neutrons) to specific organs or tissues are determined for individuals based on employment records, bioassays, and occupational radiation monitoring such as individual film badges, and dose-response analyses conducted to learn of any radiation associations (36,38–41). The approaches set out herein have been used effectively in radiation epidemiology for nearly 50 years, and health studies have improved our understanding of cancer risks following radiation as well as provided the basis for radiation protection standards for workers and the public (7,42–44).

Human subjects research approval was received from the Vanderbilt University Institutional Review Board and the Oak Ridge Sitewide Institutional Review Board. The aims specific to the MPS-CMS data linkage are covered under the strict terms of a CMS data use agreement (DUA), RSCH-2022-57727, that governs the usage, access to, and data management of the CMS linked files. The DUA requesting organization is the National Council on Radiation Protection and Measurements (NCRP). Vanderbilt University Medical Center serves as the data custodian institution. All data are stored and analyzed on secure Vanderbilt systems, and collaborators from other institutions who receive approval to access the data under the terms of the DUA do so through virtual private network (VPN) access to the Vanderbilt system.

The CMS Medicare databases capture all enrollment, medical claims, and nursing home and home health assessments for all Medicare-enrolled persons beginning in 1999 (our current data linkage includes claims data through 2021 with subsequent linkages planned for the future). Because these claims and assessments must report diagnoses, they provide a national source for identifying non-fatal and nonmalignant conditions that do not readily exist elsewhere. Data can be linked at the individual level and over time, allowing for a comprehensive look at a person’s health care utilization. The MPS links the CMS data to employment, vital status and dosimetry data of radiation workers. MPS cohorts of radiation workers come from a variety of occupations and begin as far back as 1913, i.e., radium dial painters (45). The initial linkage request consisted of 719,441workers who were alive in 1999. MPS will follow individual workers over time as they continue as Medicare beneficiaries, become Medicare eligible because of age, or were treated in Medicare/Medicaid certified nursing homes until death or the end of the study period to capture a full range of possible long-term effects of radiation exposure coupled with confounding and co-factor information. The CMS files provide a broad range of diverse health related information including incident diagnoses (e.g., for dementia and Parkinson’s disease), prescription drug use (e.g., Levodopa for Parkinson’s disease), imaging procedures (e.g., beta-amyloid PET scans for Alzheimer’s disease), confounding factors (e.g., smoking, BMI), demographic characteristics (e.g., socioeconomic status), and cognitive and physical function (e.g., standardized Cognitive Function Scale scores, Activities of Daily Living scores) and health care utilization (46).

Radiation dose to the brain and other organs for each individual has already been or will soon be estimated based on a comprehensive program of dose reconstruction that includes radiation monitors such as film badges to measure external radiation and, when available, biological samples, such as urine samples, to estimate intakes of radionuclides during employment (36). The key analyses will contrast the various outcome measures such as dementia and cognitive function by radiation dose to brain to learn if there are any associations (i.e., whether the occurrence of these outcomes or the measures of cognitive function are related to the level of cumulative radiation dose received gradually during employment in years past). The linkages with CMS databases will identify multiple health conditions including, cancer, dementia, Alzheimer’s disease, Parkinson’s disease, anxiety and behavioral disorders, cardiovascular disease, and cerebrovascular disease. Recently, there has been some discussion that Parkinson’s disease may originate with or be associated with the gut rather than or in conjunction with the brain, and we plan to evaluate dose-response with respect to colon dose as well (47).

In addition to insurance claims and enrollment data, the CMS data include federally mandated assessments for individuals receiving post-acute and long-term care from nursing facilities and home health agencies. The Minimum Data Set (MDS) contains assessments for all residents in Medicare and or Medicaid certified nursing facilities, regardless of their individual payor status or their age. Similarly, the Outcome and Assessment Information Set (OASIS) data (48) contain assessments for individuals served by Medicare and or Medicaid certified home health agencies (49–52).

The MDS and OASIS data include comprehensive assessments for each individual at regular intervals to help nursing home and home health agency staff identify needs and develop care plans. The MDS, for example, includes over 400 items to assess residents’ physical functioning, communication/hearing problems, behavioral symptoms, mood, care preferences, diagnoses, medication use, and, important to the MPS, cognitive function (50,51,53). MDS assessments are conducted by trained, licensed health professionals upon admission, quarterly, annually, upon discharge, and otherwise whenever the patient experiences a significant change in his or her health status. Approximately 17% of MDS-covered residents are under the age of 65 as Medicare coverage is not a requirement for nursing home care (54). The purpose of these linkages is to examine cognitive functional scores and other indicators for diagnoses and cognitive health outcomes following linkages with the MPS worker and veteran data with a particular interest in workers with radiation dose to brain tissue (4,8,20,24,26,38,39). Prescription drug use for the treatment of depression (e.g., Escitalopram (Lexapro)), dementia, Parkinson’s disease and related disorders were obtained via linkages with the CMS Medicare Part D files and are being incorporated into the analyses of relevant outcomes. A description of CMS data files and key variables for the MPS study is provided in Supplemental Table 1.

Diagnosis codes within CMS claims files can be used to identify chronic conditions. CMS provides chronic condition flags for as many as 70 conditions beginning in 1999 through their Chronic Conditions Data Warehouse (CCW). These condition flags are available annually for each Medicare beneficiary with fee-for-service claims and are derived from algorithms that assess the presence of specific diagnoses and procedure codes across Medicare fee-for-service claims for as much as a 3-year look-back period. CMS developed the variables from algorithms validated from the research literature and criteria used by other federal resources as well as CMS-contracted expert panels who refined and enhanced the algorithms (48). The chronic condition flags that are available from CCW in the CMS data are listed in Supplemental Tables 2 and 3. There are two available CCW files, one that covers claims data from 1999–2019 and an updated dataset that covers from 2017 onward (with several overlapping years for both files). It is notable that several flags relate directly to important study outcomes, i.e., cognitive impairment conditions and neurodegenerative diseases (i.e., Alzheimer’s disease, Parkinson’s disease, dementia, depression, stroke, anxiety disorders, depressive disorders, epilepsy, intellectual disabilities, traumatic brain injury). In addition, there are several flags in the CCW data that relate to potential confounders in epidemiologic health studies (e.g., tobacco use, alcohol use, diabetes, obesity, hyperlipidemia, hypertension, opioid use disorder). Similar algorithms for the chronic conditions will also be run against the claims files for prior years where appropriate (e.g., Parkinson’s disease).

All residents, regardless of payer status or age, who enter a Medicare and/or Medicaid-certified skilled nursing facility have an MDS assessment at admission. Prior to October 2010, patients’ cognitive function was assessed with the MDS 2.0-derived cognitive performance scale (CPS2.0) (55). The CPS2.0 score is calculated using a complex algorithm that assigns residents a score between 0–6 based on five MDS items (i.e., daily decision making, eating self-performance, ability to make oneself understood, short-term memory, whether or not the resident is comatose). The CPS2.0 scale corresponds closely with scores generated by the Mini-Mental State Examination and the Test for Severe Impairment, nursing judgments of disorientation, and neurological diagnoses of Alzheimer’s disease and other dementias (55). The MDS 3.0, introduced in October 2010 changed the calculation substantially by including the Brief Interview for Mental Status (BIMS) to assess residents’ cognitive function (56–58). For residents unable to complete the BIMS, assessors complete two items: short term memory and cognitive skills for daily decision making that enable calculation of the CPS3.0. The MDS 3.0 cognitive function scale (CFS) is a four-point scale that integrates the BIMS and the CPS3.0 (53,55,57). The CFS has high levels of construct validity (53). To reconcile the change in cognitive scoring over time, residents with a CPS2.0 score greater than 2 and a CFS score greater than 1 will be categorized as “cognitively impaired”; and residents with a CPS2.0 score of 5–6 or a CFS score of 4 categorized as “severely impaired.” These available quantitative measures of cognition among radiation workers with and without radiation dose to brain tissue will provide a powerful and unbiased assessment of whether low-dose rate exposures can cause clinically recognized memory loss and cognitive decline.

To assess cognitive function among patients receiving health care at home, items M0610 in OASIS A, B, and B1 and M1740 in OASIS C, C1, and D are used. While M0610 and M1740 are named differently, they assess the same factors over time. Specifically, assessors evaluate behaviors and or cognitive, behavioral and psychiatric symptoms that are demonstrated at least once a week (either reported or observed). These include: 1) memory deficit; 2) impaired decision-making; 3) verbal disruption; 4) physical aggression; 5) disruptive, infantile, or socially inappropriate behavior; or 6) delusional, hallucinatory, or paranoid behavior. This concludes the descriptions of the comprehensive health care data that are available, and the extensive listings of outcomes and diagnoses related to cognitive disfunction and neurological and behavioral maladies. The wealth of information available on neurological outcomes and cognitive scores, coupled with radiation dose assessment, will lead to an evaluation that hitherto has never been attempted on such a large scale and on such a unique and large population.

Estimation of absorbed radiation dose to brain tissue and other organs are performed as described in NCRP Report No. 178 and other reports (36,38–40,59,60). The absorbed radiation dose to brain tissue from high-LET radiation is calculated using new and modified approaches developed by NCRP in Commentary No. 31 (26,61).

The overarching goal for dose reconstruction for neurobehavioral analysis is to quantify occupational brain doses from external radiation and internal radionuclides. Brain doses and associated uncertainties are based on the guidance published by NCRP in Report No. 178 (36). Under this approach, the brain is assumed to be the organ of interest for neurodegenerative diseases, depression and cognitive dysfunction. Accordingly, brain doses for each cohort worker is reconstructed on an annual basis beginning from the first occupational radiation exposure until death, incidence of an outcome of interest, or end of follow-up. Initially, the total absorbed organ dose will be partitioned into low- and high-LET components without any assumptions about the biological effectiveness of any radiation types. Regarding external irradiation, the personal dose equivalent H_p(10) for workers is often available based on federally-required personnel monitoring. A substantial part of the dose reconstruction and uncertainty estimation is devoted to estimating the organ doses based on the records of H_p(10) as outlined by NCRP (36) and is considered in all aspects of dosimetry such as missing dose, multi-facility career dose, irradiation geometry, radiation spectra associated with likely types of external irradiation scenarios, and partial organ irradiation. External exposures include both photon and neutron exposures.

The MPS benefits from collaboration with the United States Transuranium and Uranium Registries (USTUR), a research program established as the National Plutonium Registry in 1968 by the Atomic Energy Commission (62). Designed as a program to improve radiation protection of nuclear workers, the USTUR studies the biokinetics and internal dosimetry of actinides (uranium, plutonium, and americium) in occupationally exposed workers (63–65). These workers voluntarily donated portions of their bodies, or their whole bodies, for scientific use posthumously. Today, the USTUR is one of the few programs worldwide that can comprehensively study biokinetics and dosimetry of internally deposited actinides. Currently, the USTUR and associated National Human Radiobiology Tissue Repository (NHRTR) hold data and tissue samples from 48 whole-body and 317 partial-body donors. Since 1993, USTUR/NHRTR has housed collections of tissue materials and records obtained from the Radium Worker study conducted at Argonne National Laboratory (65,66). The USTUR/NHRTR is a unique scientific resource (https://ustur.wsu.edu) that includes information on brain activity deposition, and biokinetics valuable for improving and refining the estimates of cognition-related tissue and organ doses (61).

Epicure, a software package frequently employed for radiation epidemiology studies, is functionally capable of analyzing datasets with millions of rows of data; however, the computational demand for MPS is so large that alternative software is being developed (67). With the very large size of the input data for a million or more workers, each worker with annual radiation doses for multiple years, outcomes, confounding and co-factor data, other demographic and occupational data, and information for uncertainty analyses, new software is being developed to handle fitting risk models and perform uncertainty analyses. This need for new software is being met by creating a library in open-source R statistical programming language (68) that will be optimized to handle input on the scale of over 20 million records and upwards of a billion data variables. The new software, called Colossus (69,70), can fit Cox regression models and Poisson regression models, similar to the PEANUTS and AMFIT programs within Epicure, respectively (71–73).

Work to develop such tools to run partial likelihood methods (such as Cox regression) and Poisson regression is underway using code written for the open-source R statistical programming language and incorporating novel methods for speeding up the model fitting runtime on multi core, modern fast computers. The basis of the software approach is that R has the desired statistical tools and C++ can load and operate on large datasets. The software uses R for finding summary statistics and uses C++ to perform more intensive computations. The survival library (74,75) and stats library in R can perform Cox proportional hazards and Poisson regression modeling effectively on data with millions of person-years. Comparison with smaller datasets can be done with Statistical Analysis System (SAS) and Epicure. Development of R software is being accomplished by using the Rcpp library to transfer the data into C++ and then using the RcppEigen library to integrate the capabilities of the Eigen library and treat the calculation as a series of optimized matrix operations. Validated and documented statistical software packages will be deposited in public repositories (i.e., data at the individual level of one person year at risk) (76).

Data analysis will be based on fitting risk models to combined datasets, to estimate radiation-related risks for specific outcomes, either as a function of the relevant organ/tissue dose, e.g., brain, or risk per unit dose. The risk models will initially be based on linear dose-response models in which the radiation risk either adds to the baseline risk (excess absolute risk) or multiplies the baseline risk (excess relative risk). A further step in the software development will be to consider Multi Model Inference, which allows risk estimates to be based on several plausible dose-response models rather than relying on one single model of choice, the most common being the linear non-threshold (LNT) model (44, 77, 78). Both Cox and Poisson risk analysis methods will be included to fully characterize both the baseline risk and the excess risk due to radiation. Further, ensemble models (79, 80) and other methods (81) for addressing uncertainties will be considered.

NASA evaluates nervous system risks in two categories: acute central nervous system (CNS) risk that could lead to in-mission performance decrement with a risk of mission loss and or crew loss, and late CNS risk such as the development of neurodegenerative diseases post mission. NASA uses data from computational models, preclinical animal and cell models, and human data to characterize, manage, and quantitatively assess radiation risks (32,86–88). There are several limitations of the existing data for evaluating neurodegenerative diseases within spaceflight crews, specifically the small number of astronauts who have accumulated relatively high doses and their short follow-up for outcome evaluation (86–89). Furthermore, future missions that operate outside of LEO will experience prolonged exposures with increased cumulative doses from the complex radiation environment of GCR, including the energetic high-LET heavy ions found therein (79). Therefore, the data from the present study of terrestrial workers will provide important information from human cohorts to inform the NASA risk models for space radiation and approaches to risk mitigation, especially for long duration missions including trips to the Moon and Mars (2). Mars mission-related doses are expected to be of the order of 300–450 mGy (870–1200 mSv) (1,90–93) and are consistent with the radiation exposures in the MPS (2), in which exposures occurred over many years and not over minutes or weeks as in previous animal experiments or in the Life Span Study (LSS) of the atomic bomb survivors. Information from high-LET and low-LET brain exposures in humans, in combination with other computational and preclinical models, will aid in the NASA assessment and management of radiation risks (26,94–96).

However, it is recognized that no human terrestrial radiation exposures are ideal analogues to space radiation exposures as the doses, dose-rates, and radiation quality characteristics differ. No human terrestrial exposure can replicate the scenario of space radiation where heavy ions are traveling near the speed of light, and no terrestrial environment is comparable to a space station or vehicle (4). Nonetheless, when assessing energy deposition patterns in cell nuclei, there are overlaps between GCR versus terrestrial alpha emitting particles showing potential relevance for high-LET doses (95–96) (Fig. 1).

NASA has supported the development of tests that assess multiple domains of neurocognitive functions applicable to space flight including assessment of working memory, stability of sustained attention, and risk decision making. The Cognition Test Battery (COGNITION) has replaced the WinSCAT cognitive performance battery of tests (97–99). COGNITION uses multiple, brief, neurobehaviorally orthogonal tests, is neuroimaging based, and has been validated and normed on astronauts and other high performers (100). It is primarily used to assess changes in cognitive function (pre-mission, in-mission, and post-mission) and was used in the NASA Twin Study (101–102). However, long-term cognitive impairment, and risk for neurodegenerative diseases in astronauts, remains understudied (100). Using the CMS data that contain diagnoses and assessments of cognition – for MPS radiation workers receiving care from nursing facilities and home health agencies – combined with data detailing chronic radiation exposure in a terrestrial occupational environment will provide baseline data to inform risk assessments for space travel. Estimates of health effects related to cognition, behavioral changes, dementia and depression will be possible, and this human experience should help decision-makers in assessing radiation risks and setting radiation protection guidance.

Medicare claims data are frequently used to identify prevalence and incidence of health conditions among older adults. For example, by comparing claim-based identification of Alzheimer’s disease and related dementia to results from cognitive assessment tests, researchers have established that Medicare claims data for adults aged 65 and over can identify most cases of Alzheimer’s disease or dementia but is less successful at identifying mild cases of dementia and is more likely to identify dementia among White adults than Black or Hispanic adults (103–104). Claims-based identification of Alzheimer’s disease and related dementia has been used to explore differences in prevalence or incidence related to race and ethnicity, lifestyle factors, and statin use, among others (84,105, 106). Thus, procedures are in place to identify large numbers of workers with incident cases of Alzheimer’s disease and related dementia, along with important demographic and clinical co-factor data, and evaluate associations with radiation dose to brain. Alzheimer’s disease is of importance because of animal models suggesting an increase in cognitive impairment and αβ plaque accumulation following galactic cosmic radiation (107).

Quantification and propagation of uncertainties is crucial (1) for understanding the precision in outcome risk estimates obtained through analysis of the CMS datasets and (2) for later use when estimating risks and the associated uncertainty for individuals exposed to an annual series of doses (80, 86). Typically, excess risk parameterization includes an implicit assumption that the subject-specific covariates, including doses, are exact, and so the calculated confidence intervals do not include uncertainty contributions from these variables. While dosimetric uncertainty is always important to characterize, it is especially crucial when analyzing longitudinal series of low annual doses such as those considered in the MPS cohorts, where the uncertainty associated with annual doses may vary as a function of any number of dosimetric variables, including time.

The MPS is exploring ways to implement uncertainty approaches similar to those described in Federal Guidance Report (FGR) No. 13 (108) that tabulates cancer risk coefficients for internal or external exposure of the U.S. population to each of more than 700 radionuclides in environmental media such as air and drinking water. FGR 13 characterizes the expected reliability of individual risk coefficients in terms of broad, semi-quantitative uncertainty categories identified by letters A-E (109). Category A represented the most narrowly determined risk coefficients, Category E represented the least well characterized coefficients, and Categories B, C, and D represented intermediate, increasing levels of uncertainty. The latest available biokinetic models as described in NCRP Commentary No. 31 (26) are being used to predict uptake and retention of absorbed activity by the brain. The uncertainty in brain dose estimates for intake of any of the dosimetrically dominant radionuclides encountered so far at MPS sites (e.g., U or Pu isotopes, Ra-226, Po-210, and Am-241) are usually assigned to Category B or lower categories.

Recently, the NCRP published Commentary 34 (81) on dose uncertainties in radiation epidemiologic models which will be carefully considered. One possible method of dosimetric uncertainty propagation being considered is to use Monte Carlo sampling to generate alternative, plausible realizations of annual dose series for each subject in the study using known or estimated probability density functions for each annual dose, and then reanalyze the entire dataset to generate the corresponding excess hazard function and associated confidence intervals for the corresponding realization. The Monte Carlo sampling process is repeated until statistical convergence of the mean excess hazard function is achieved and the construction of a joint probability density function for all excess hazard function model parameters is enabled. This method for dosimetric uncertainty propagation is computationally extensive and the regression must be performed hundreds or possibly thousands of times. Use of analysis code purpose-built for processing large datasets that employs extensive parallelization could make such an approach tractable. Model form uncertainty has historically not been considered in radiation epidemiology studies but could be included in a manner similar to dosimetric uncertainty. Here, probability density functions for the relative contribution to a linear combination of each of the proposed model forms could be defined, possibly informed by analyses of other cohorts and expert judgment (110–1127). A large number of realizations of the overall dose response form could be generated and used for regression. Propagation of covariate uncertainty (i.e., dosimetric uncertainty) and model form uncertainty could occur simultaneously since these are independent contributors to the overall uncertainty in excess hazard function.