7th Asian and Oceanic Congress on Radiation Protection, Beijing, China, September 8–12, 2026
Misclassification of causes of death in radiation epidemiology: Validation, simulation, and correction
Xirui Liu (USTUR), Stacey L. McComish (USTUR), Joey Y. Zhou (DOE), Sergey Y. Tolmachev (USTUR)
Introduction: Epidemiological studies of populations exposed to radiation provide the scientific foundation for radiation risk assessments and the establishment of radiation protection standards. Dose-response epidemiological studies typically rely on the underlying cause of death listed on death certificates to determine health outcomes. However, a study by the United States Transuranium and Uranium Registries (USTUR) of former nuclear workers found that approximately 25% of the donors’ death diagnoses on death certificates did not match the corresponding autopsy report [1]. Therefore, this study focused on such misclassification, with three main objectives: to (1) calculate the actual misclassification rate of underlying cause of deaths across different disease categories among USTUR Registrants; (2) quantify, through simulation experiments, the impact of these misclassifications on the statistical significance of dose-response model conclusions; and (3) apply statistical correction method for outcome misclassification errors published by Rogan and Gladen [3], and evaluate its effectiveness and potential risks.
Methodology: This study consisted of three phases. Phase 1 was an observational validation study in which the cause of death from autopsy reports and death certificates of 268 deceased former nuclear workers were compared at the level of the primary ICD-10 (International Classification of Diseases, 10th Revision) disease chapter. For overall and each disease category, the match rate, the over misclassification rate (false positive rate), and the under misclassification rate (false negative rate) were calculated, and logistic regression was used to identify factors that influenced match accuracy [1]. Phase 2 was a simulation study that first generated a reference non-misclassified dataset which paired radiation exposure doses with cancer deaths and was selected to have a borderline non-significant p-value. Cancer and non-cancer outcomes were then randomly reassigned at over- and under-misclassification rates ranging from 0% to 30%, with 20,000 iterations per rate combination. For each iteration, the misclassified odds ratio and corresponding p-value were calculated and compared to those of the reference dataset. For each rate combination, the percentage of misclassified odds ratios that moved away from the null with p-values that shifted from non-significant to significant was calculated to reflect the impact on conclusions [2]. The influence of the disease baseline and the significance threshold on this conclusion-change percentage was then further analyzed. Phase 3 used preset underlying odds ratios ranging from 1 to 4.5 to generate 10,000 population, misclassified, and corrected datasets. The misclassified datasets were corrected by fitting each dataset with a logistic regression model and determining the probability of disease as a function of dose. These probabilities were then corrected and used to generate the corrected datasets.
Results: Among the 268 USTUR Registrants, the overall misclassification rate was 25.4%, and match status varied substantially by disease category. The reliability of diagnoses was relatively high for neoplasms category, which had 4.3% over misclassification rate and 13.3% under misclassification rate. Respiratory diseases showed the highest under misclassification rate at 47.4%, while circulatory diseases showed the highest over misclassification rate at 12.2%. The odds of a match were 3.4 times higher when death certificates were completed using autopsy findings, and 2.8 times higher when clinical history was mentioned in the autopsy reports. Phase 2 simulation results revealed that misclassification could artificially alter the statistical conclusion between 9.4% and 42.8% of the time in borderline scenarios, and lower misclassification rates were associated with a higher percentage of altered conclusions. For over misclassification rates greater than 5%, increasing the baseline disease rate increased the likelihood of a conclusion change. Raising the significance threshold, for example from p < 0.05 to p < 0.10, increased the percentage of affected conclusions across all rate combinations, indicating a greater overall potential impact on conclusions when other conditions were held constant. In Phase 3, the correction method successfully recovered the underlying odds ratio associations, but the corrected odds ratio distributions became wider than the misclassified odds ratio distributions, leading to inflated false positive and false negative rates. For example, the false positive rate was simulated by generating datasets with an underlying odds ratio of 1 and determining the proportion with significant p-values. When this was done, the false positive rate was 5% among both population and misclassified datasets, but increased to 21% as a consequence of the correction method. Similarly, false negative rates were simulated using underlying odds ratios that were greater than 1, and determining the proportion of non-significant p-values.
Conclusion: Misclassification of causes of death is a non-negligible challenge in radiation epidemiological modeling. The validation data confirmed that approximately 25% of death certificates may be misclassified regarding the underlying cause of death, particularly for respiratory and circulatory diseases. Simulation results suggest that misclassification errors may artificially produce statistically significant dose-response associations, especially in studies that adopt high significance thresholds or that examine diseases with higher baseline prevalence. Although the Rogan-Gladen correction method successfully recovered the underlying odds ratio associations, the wider corrected odds ratio distributions inflated both false positive and false negative rates. Therefore, making effective use of clinical records and autopsy findings remains the most reliable strategy for enhancing data quality in risk assessments. These findings serve as a reminder to remain cautious when interpreting radiation epidemiological studies that rely only on death certificates, and they provide guidance on key considerations for establishing future radiation protection safety guidelines. [USTUR-0744-26A]
Keywords: death certificate misclassification, radiation epidemiology, dose-response
References:
[1] McComish, S. L., Liu, X., Martinez, F. T., Zhou, J. Y., & Tolmachev, S. Y. (2024). Misclassification of causes of death among a small all-autopsied group of former nuclear workers: Death certificates vs. autopsy reports. PloS one, 19(5), e0302069.
[2] Liu, X., McComish, S. L., Howard, S. C., Zhou, J. Y., & Tolmachev, S. Y. (2026). Inference under outcome misclassification in health risk models using a simulation study with a validation dataset. Scientific reports, 16(1), 11981.
[3] Rogan, W. J., & Gladen, B. (1978). Estimating prevalence from the results of a screening test. American journal of epidemiology, 107(1), 71‒76.
USTUR human data: Enhancing radiation epidemiology through dose uncertainty and outcome misclassification assessment
Joey Y. Zhou (DOE), Maia Avtandilashvili (USTUR), Xirui Liu (USTUR), Stacey L. McComish (USTUR), Sergey Y. Tolmachev (USTUR)
Introduction: The United States Transuranium and Uranium Registries (USTUR) is a unique and invaluable human-data
resource for advancing radiation epidemiology. In combination with detailed work history, bioassay data, and radiation
exposure records, the Registries collects samples at autopsies from volunteer tissue-donors with known history of exposure
to actinide elements.
Methodology: Post-mortem actinide organ activity measurements allow to validate and improve actinide biokinetic
models and provide a unique opportunity to quantify uncertainties in radiation dose assessment. Autopsy reports, available
for the entire cohort, provide a rare opportunity to assess the accuracy of death certificate data commonly used in
epidemiologic studies.
Results: Recent USTUR studies [1-3] demonstrated that: (i) biokinetic model predictions often fail to capture measured
post-mortem organ activities, which can serve as a proxy for the radiation dose, (ii) death-certificate-based causes of death
show substantial disagreement with autopsy findings, and (iii) even non-differential outcome misclassification can
meaningfully distort epidemiologic dose-response estimates. These studies underscore USTUR’s indispensable role in
strengthening the scientific foundation of radiation epidemiology.
Conclusion: By providing direct human-based validation data, the USTUR supports improved dose reconstruction, more
realistic uncertainty quantification, and critical evaluation of the influence of health outcome misclassification on
epidemiologic dose-response estimation and interpretation. [USTUR-0745-26A]
Keywords: Death certificate misclassification, radiation epidemiology, plutonium organ activity
References:
[1] McComish, S. L., Liu, X., Martinez, F. T., Zhou, J. Y., & Tolmachev, S. Y. (2024). Misclassification of causes of death among
a small all-autopsied group of former nuclear workers: Death certificates vs. autopsy reports. PloS one, 19(5), e0302069.
[2] Liu, X., McComish, S. L., Howard, S. C., Zhou, J. Y., & Tolmachev, S. Y. (2026). Inference under outcome misclassification
in health risk models using a simulation study with a validation dataset. Scientific Reports, 16(1), 11981.
[3] Avtandilashvili, M., Sefl, M., Zhou, J. Y., & Tolmachev, S.Y. (2025). Validation of Bayesian modeling approach of uncertainty
in organ doses using post-mortem measurements. Scientific Reports 15(1), 20476.
AI-Agent Assisted USTUR Internal Dosimetry Research: A Human-AI Coworker Workflow
Joey Y. Zhou (DOE), Maia Avtandilashvili (USTUR), Sergey Y. Tolmachev (USTUR)
Introduction: The United States Transuranium and Uranium Registries (USTUR) is a unique resource for developing
and testing human biokinetic models for internally deposited actinides. The Registries provide a rare opportunity to
directly compare model-predicted actinide organ activities with those radiochemically measured post-mortem
(autopsy data). Artificial intelligence (AI) is increasingly used across scientific disciplines to accelerate data analysis,
support complex modeling, and enhance research workflows. This work evaluated the use of an AI coding agent as
a supervised research coworker for dosimetric application of USTUR data.
Methodology: A project workspace included a paper on 20 cases from two worksites [1], recommended biokinetic
models, in-house Python code for plutonium biokinetic modeling, and USTUR case data sets. The AI agent assisted
with code inspection, archive searches, data extraction, model execution, posterior summaries, report generation,
and iterative scientific interpretation. The workflow used an independent Markov Chain Monte Carlo (MCMC)
implementation to reproduce the original 20-case analysis based on an assumption of chronic inhalation intake and
use of urine bioassay data only. The analysis was then extended to 35 cases with different intake scenarios and data
configurations: (i) chronic inhalation – urine only; (ii), chronic+acute inhalation ‒ urine only; (iii) chronic inhalation
‒ urine +autopsy data; and (iv) chronic+acute inhalation ‒ urine only and autopsy data.
Results: The expanded analyses confirmed that using autopsy data can reasonably well predict respiratory tract and
systemic plutonium activities, whereas urine only analyses provide weak and sometimes misleading respiratory tract
inference, as nominal 95% coverage can be driven by very wide posterior intervals. Sensitivity analyses of intake
scenario assumptions and additional respiratory tract dissolution and deposition parameters did not show significant
effects on model predictions.
Conclusion: AI agents, when used as transparent and supervised scientific coworkers, can accelerate reproducible internal
dosimetry research. [USTUR-0746-26A]
Keywords: Plutonium; internal dosimetry; bioassay; Bayesian analysis; artificial intelligence agent
References:
[1] Avtandilashvili, M., Sefl, M., Zhou, J. Y., & Tolmachev, S.Y. (2025). Validation of Bayesian modeling approach of
uncertainty in organ doses using post-mortem measurements. Scientific Reports 15(1), 20476.