Overcoming Clinical Validation Challenges for an In Vitro Diagnostic Biodosimetry Test

Abstract

Introduction

In the aftermath of a nuclear detonation or other mass-casualty radiological event, thousands to millions of people may be exposed to high levels of radiation. Effective triage tools are needed to identify those with potentially life-threatening exposure requiring emergency medical care. To ensure these tools will be available to support a mass-casualty radiation event, US Food and Drug Administration (FDA) authorization is required. However, currently, no FDA-authorized biodosimetry tests are available to support a mass-casualty radiological event, due largely to the significant challenges in developing regulatory strategies for biodosimetry tests.

Materials and Methods

We have developed a tool, called the CellRADx Software Application, to help address this mass-casualty triage need. CellRADx utilizes complete blood count (CBC) results from a single test, along with other inputs such as age and time post-exposure, to determine if an individual has absorbed more than 2 Gy of ionizing radiation. As part of the development of CellRADx, we have created a novel approach to facilitate the complete validation of the software device in compliance with FDA’s medical device regulations. This novel approach employs a combination of real-world CBC data, an in silico radiation-induced cell depletion model, and nuclear event dose distribution models to validate clinical performance.

Results

Performance of CellRADx across the general population has been demonstrated with high sensitivity (92%) and high specificity (78%). Additional subpopulations with potentially confounding conditions were also tested using this approach and have demonstrated similar performance results.

Conclusion

The ability to validate a biodosimetry device using a method that can support FDA authorization is a key enabling development that will facilitate implementation of tools that can improve preparedness for a mass-casualty nuclear event and help save many lives.

INTRODUCTION

The chaotic aftermath of a nuclear detonation in a major US city will require an unprecedented, integrated coordination of a vast number of resources and actions to achieve an effective response. A significant and unique challenge will be assessing radiation exposure for hundreds of thousands to millions of potentially exposed people to identify those that require immediate medical treatment, and just as importantly, identify those at low risk that should be medically cleared so that sparse and precious resources can be dedicated to those most in need of them. To accomplish this goal, emergency responders will need access to FDA-authorized biodosimetry triage tools appropriate for this challenge.

To address this need, we have developed the CellRADx software application with an embedded algorithm to identify individuals who were likely exposed to more than 2 Gy of ionizing radiation. A user enters the CBC results from a single sample from a patient into the application, along with some additional supporting information, such as the patient’s age, when the CBC blood sample was taken, and when the suspected exposure occurred (Figure S1). The CellRADx software determines a positive result (exposure likely greater than 2 Gy) or negative result (exposure likely less than 2 Gy). This approach has several key benefits, including that (1) it only requires CBC results at a single time point, substantially simplifying logistics during the chaotic response to a nuclear disaster, (2) the software can use CBC inputs from a variety of existing, widely-available, FDA-cleared CBC devices, eliminating the need to stockpile dedicated test equipment or consumables, and (3) its open architecture facilitates testing in a variety of locations wherever CBC testing is available, such as urgent care clinics, doctor’s offices, hospitals, and temporary medical facilities, adding significant operational flexibility. We envision multiple methods to rapidly deploy the CellRADx application in resource limited environments, including web-based hosting accessible from virtually any device with an internet connection, standalone versions of the app for iOS and Android devices, copies stored on flash drives in the Strategic National Stockpile, and direct integration with widely installed CBC devices and hospital EHR/EMR systems.

However, a particular challenge shared across all biodosimetry tests, including CellRADx, is how to perform clinical validation to support regulatory approval with a lack of access to a human cohort representative of the intended use population. Clinical samples from patients exposed to radiation resulting from a nuclear detonation do not exist or cannot be prospectively generated by testing subjects intentionally exposed to life-threatening doses of radiation. Patients receiving radiation therapy (RT) for various cancers and other conditions represent a possible surrogate population, but their treatment regimens typically consist of fractionated doses administered in a series over time. Although the accumulated fractions can represent a large total-body dose of radiation, the exposure methods do not align with the intended use scenario of a single acute exposure, complicating the ability to validate performance over time post-exposure, since biodosimetry biomarkers typically vary in magnitude across both absorbed dose and time post-exposure. Additionally, these RT patients’ underlying diseases (e.g., leukemias) and concomitant treatments (e.g., chemotherapy) are not representative of the general public and can often impact the underlying biomarkers of biodosimetry tests, confound results, and greatly limit the usefulness of this population.

The use of animal models, such as non-human primates (NHPs), is another option as a model for the intended use population, but unlike for drugs and biologics, the Animal Rule¹ does not apply to in vitro diagnostic (IVD) tests. Further, although the human hematopoietic system is similar to various other animal models, with NHPs perhaps offering the closest match, there are still pertinent differences that result in a need to perform complex bridging studies, adding uncertainty to results and raising questions about model sufficiency from FDA.

More broadly, the use of RT patients and/or animal models as surrogates or model populations does not address the need to demonstrate device performance in various subpopulations that may have underlying conditions or concurrent injuries (e.g., thermal burns, physical trauma) that can affect test performance. Factors such as age, sex, obesity, diabetes, various immune diseases, active viral and bacterial infections, and common medications can all impact various biodosimetry biomarkers. For regulatory authorization, it is necessary to validate device performance in the intended use population, which includes these subpopulations.

To address these limitations, we have developed a novel approach to clinically validate the CellRADx software and its underlying analysis algorithm. Using CBC data collected from various unirradiated populations and published computational models of blood cell count changes over time after radiation exposure, we can generate simulated CBC data at any time post-exposure and relevant dose level. This process can be performed for virtually any subpopulation of interest using CBC data from representative (but unirradiated) subjects. Next, we apply this collection of in silico data (representing irradiated conditions at various times and doses) to the expected distribution of radiation exposure levels based on published nuclear-blast modeling. This effectively simulates the prevalence of expected exposure amounts after a nuclear detonation, as we can properly account for the percentage of the population who received 0 Gy, 1 Gy, 2 Gy, and so forth. Lastly, we can process this collection of data through the CellRADx software, just as if we were prospectively testing subjects after a nuclear blast. Thus, the performance of the device can be assessed, with various metrics (e.g., sensitivity, specificity) calculated, and the results can be used to support a regulatory submission to the FDA for marketing authorization.

METHODS

Collecting Available Accident Data and Unirradiated Data

The intended use (IU) population for biodosimetry devices (i.e., individuals who are potentially exposed to a significant, acute external dose of ionizing radiation) is not widely available and cannot be created through traditional prospective studies. Retrospective human, real-world data (RWD) are available in very limited quantities from historical radiation accidents, which serve as the best representation of the IU population. We have acquired access to data from these populations for development and training of the CellRADx Software and sequestered portions of these datasets for exclusive use in future clinical validation. Radiation accident victim data used for development and training have been primarily acquired from the Armed Forces Radiobiology Research Institute (AFRRI) and the Radiation Emergency Assistance Center/Training Site (REAC/TS), as well as from publications by the International Atomic Energy Agency (IAEA). These curated data include information specific to each victim, including the equivalent total-body dose estimate, the CBC measurands observed, and sample times post-exposure. Where possible, we have collected information on the dosimetry methods used to estimate ground truth dose, such as dose-reconstruction with cytogenetic methods, for example, the Dicentric Chromosome Assay. Where available, patient demographic information was also captured, including age and sex.

In support of validation of the CellRADx software, we have sought access through ongoing collaboration with the Bundeswehr Institute of Radiobiology (Munich, Germany) to an additional source of radiation accident victim data in the SEARCH (System for Evaluation and Archiving of Radiation accidents based on Case Histories) database.² This database contains a large selection of cases that are distinct from those used for algorithm development. We intend to keep these data sequestered for use solely in clinical validation of the CellRADx software which has not yet been performed.

Although RWD from radiation accident victims are extremely valuable, they are insufficient in quantity and diversity (demographics, dose levels, time post-exposure, etc.) to fully support device validation. Therefore, we have also obtained retrospective data for additional populations representing the unirradiated population. In particular, we extracted data from the National Health and Nutrition Examination Survey (NHANES), a public database managed by the Centers for Disease Control and Prevention that contains health data on roughly 100,000 human participants representative of the United States general population.³ This database provides CBC results, in addition to demographic details (age, sex, ethnicity, etc.) and health information (e.g., pre-existing medical conditions, common medications). We extracted data from approximately 40,000 subjects for use in development and sequestered a roughly equivalent amount for later use in clinical validation. This data is used in both its native (unirradiated) form, as well as an input to the cell depletion model (described in subsequent sections herein) to generate in silico data representing an irradiated population.

Cell Depletion Models

We have utilized computational models of human hematopoiesis to generate surrogate data for irradiated populations. In basic terms, the hematopoietic cell depletion model is used as a method to convert blood cell counts from unirradiated subjects into predicted blood cell counts for irradiated subjects at any given time between 1 and 7 days post-exposure for radiation doses between 0 and 10 Gy. Additionally, the FDA released guidance in 2023 to clarify expectations for device developers intending to use modeled data in support of device validation.⁴ We have applied this Guidance to use in silico derived cell counts to measure clinical performance in lieu of a prospective clinical trial.

The model is a multi-compartment, deterministic model based on scientific research of the functions of the hematopoietic system and the effects of ionizing radiation. We selected multi-compartment models based on systems of differential equations to model various steps in the life cycle of particular cell types, as models of this format have been employed in numerous studies to specifically represent hematological responses to ionizing radiation.^5-10 As shown in Figure 1, the basic structure of these models is to represent the maturation of a given blood cell type through multiple stages, for example, dividing precursors to mature peripheral cells, that are treated as separate “compartments.” The transition of cells through these compartments is modeled with differential equations and various parameters to represent rates, inhibitory factors, and other interactions.

Figure 1.

Basic structure of multi-compartment models used to represent the transition of a given cell type through multiple stages of maturation. (left) The model represents cell stages as “compartments” X₁, X₂, and X₃. Transition through the compartments is governed by various parameters to represent rates (B, C, F, and E), inhibition factors (I), and other interactions (Θ₁, Θ₂, Θ₃). (right) The basic format of governing differential equations is used to build the mathematical model, which includes additional parameters (G, H) to weight interactions. Adapted from Hu, Smirnova, & Cucinotta⁵ and Wentz.⁹

In general, these parameters are experimentally defined, based on prior basic research on hematopoiesis and radiation-induced cell damage. Various research groups have developed models of radiation-induced hematological changes based on this fundamental structure, although slight variations are common, for example, the format of inhibitory factors and numbers of compartments for a given cell type. We selected the specific models developed for the Defense Threat Reduction Agency (DTRA) as described in their 2014 report, DTRA-TR-14-31.9 This report was developed from prior work by Hu and Smirnova,^5,7 with minor changes by the investigators to some experimentally defined parameters and optimization of others to improve the fit of the model to data from irradiated human subjects (radiation accident victims and radiotherapy patients). More details on the exact equations and parameter values utilized can be found in this published DTRA report.

We programmed the models into a set of analysis tools using RStudio software (posit, Boston, MA, United States). These tools allowed us to input initial, unirradiated cell counts and specify an exposure dose to simulate. The tool then outputs hematology curves, representing the changes in peripheral blood cell counts over time. A given timepoint can be selected to represent a snapshot in time, and data extracted to provide a simulation of an irradiated state (Figure S2).

Under ideal conditions, evidence that this cell depletion model behaves as expected would involve individual subject comparisons, wherein RWD from irradiated human subjects are juxtaposed with modeled cell curve data generated using pre-irradiation cell counts from the same subjects as inputs. This approach is only possible if pre-irradiation CBC data are available. Unfortunately, radiation accident victims are not typically examined at times closely preceding an incident, and RT patients have abnormal cell counts even before receiving radiotherapy. However, population-based analyses are still possible since initial CBC values can be input to the model across the range of generally normal physiological values. This approach allows simulation of a range of expected values that can be directly compared to real-world radiation accident data. We generated such comparisons using inputs derived from published averages and normal ranges of counts.^11,12 There is general consistency between the modeled data and the RWD. An example comparison is provided in Figure S3.

Dose Distribution Models

The intended use population for the CellRADx software device consists of the general population in the area impacted by a large-scale radiological or nuclear incident. Therefore, a key aspect of defining the intended use population is determining the range of absorbed doses expected within that population. Consistent with CLSI EP-12,¹³ patient specimens used to measure diagnostic accuracy should include a representative population of ‘clinical states’ expected in the clinical practice. The dose of radiation received by a given individual is a key “clinical state” in the case of a biodosimeter and will depend on many factors, including but not limited to the type of detonation (ground/air), yield of the nuclear device (e.g., 1 kT, 10 kT, etc.), the distance from the epicenter, local weather conditions, and the amount of shielding from buildings. Since there are, thankfully, no recent, well-characterized instances of a nuclear detonation in a major city, this information does not readily exist and cannot be directly assessed. Therefore, we used published models of radiation exposure to estimate the expected distribution of exposure levels across the dose range in a real-world disaster scenario. Here, we summarize examples based on work published by Buddemeier¹⁴ and Knebel.¹⁵

As shown in Figure 2, Buddemeier presents modeled data from a 10kT ground burst detonation in New York City and estimates the number of casualties in several “damage zones”, that is, the Severe Damage Zone (SDZ), Moderate Damage Zone (MDZ), Light Damage Zone (LDZ) and Dangerous Fallout Zone (DFZ). Knebel presents modeled injury data from a variety of nuclear detonation scenarios including ground and air bursts of different scales, and a composite of the different models.

Figure 2.

(Top) model of 10 kt denotation in New York city. Source: B. Buddemeier presentation.¹⁴ (Bottom) graphical illustration showing decomposition of affected groups into more narrowly defined dose bins.

Both Buddemeier and Knebel estimate the number of casualties in the zones/scenarios utilizing several broad injury categories. These broad injury categories provide a very coarse distribution of radiation exposures, but do not provide sufficient resolution to inform the distribution of doses for future validation studies, particularly around the 2 Gy cutoff and in the lower doses where most of the intended use population is expected. In order to spread the number of casualties in each of these major injury categories across a finer set of dose bins to analyze diagnostic performance, we have calculated a dose distribution density–the number of people assigned to a given dose bin–and then allocated the major injury categories into the respective dose bins to approximate a smoothly decreasing dose distribution density although maintaining the approximate number of casualties in each major injury category (see Figure 2). Similar approaches have been used in other biodosimetry studies¹⁶ to estimate the most realistic distribution of doses in the intended use population, based on published disaster models.^17-19

We utilized the technique above to characterize the dose distributions associated with a variety of damage zones in the Buddemeier model and detonation scenarios in the Knebel model to represent a variety of intended use populations in nuclear incidents of different types and sizes, rather than tying performance to only a single scenario (see Table S1). As we developed the estimated dose distributions associated with the various Buddemeier damage zones and Knebel scenarios, we noted that although the magnitude of the casualties varied significantly across variations of the models, the general distribution of doses followed a similar pattern across different zones and scenarios. Since the intended use scenario may include all of these variations, we evaluated the clinical performance of the CellRADx software utilizing a range of these scenarios. For simplicity, subsequent portions of this paper describe analyses conducted with a single scenario: the Buddemeier New York City (NYC) model with three disaster zones (MDZ, LDZ, and DFZ). This template represents a reasonable worst-case scenario with a high number of affected people for a disaster in a large metropolitan area.

Integrated Clinical Performance Assessment

Clinical validation of the sensitivity and specificity of the CellRADx algorithm is based on a conventional study design for a qualitative diagnostic test. However, because of the impossibility of executing a prospective human clinical trial for the intended use, we used retrospective RWD in combination with in silico data generated using the cell depletion model in place of prospective clinical enrollment to validate clinical performance, as shown in Figure 3.

Figure 3.

Illustration of clinical validation approach for assessing sensitivity and specificity. Real world data (RWD) from unirradiated populations is used to generate surrogate data for irradiated subjects, which is then combined with RWD from radiation accident victims and data from unirradiated subjects into a virtual cohort. The number of subjects at each radiation dose level within the virtual cohort is adjusted to align with a dose distribution model and then processed using the CellRADx algorithm to assess sensitivity and specificity.

To determine an appropriate sample size, we used FDA’s Emergency Use Authorization template for diagnostic tests for COVID-19²⁰ as a guide (molecular and antigen tests for home use, November 2021), since the CellRADx regulatory pathway is likely to follow a similar EUA approach. FDA’s template calls for a lower bound (LB) of the two-sided 95% confidence interval (CI) for positive percent agreement (PPA) to be not less than 70% (10% lower than the PPA minimum of 80%). Similar performance requirements (minimum of 80% PPA or sensitivity) are also described in other EUA templates for antigen tests, such as for Monkeypox.²¹ Given an average NEG to POS ratio of 9:1 anticipated in a real-world nuclear disaster scenario (based on established dose distribution models described previously), the number of positive cases drives the overall sample size requirements because it is the significantly smaller population. In the case of CellRADx, we have established a performance target of 90% sensitivity. Applying similar LB criteria, an 80% LB of the 95% CI of the 90% sensitivity target requires 62 POS samples. Using the NEG to POS ratio of 9:1 results in the need for a corresponding 558 NEG samples. Therefore, approximately 600-650 total datapoints are required to assess sensitivity and specificity. Further, these datapoints are distributed across the target time range of 1 to 7 days post-exposure, matching the device’s intended use.

RESULTS

Overall Performance with General Population

Using the approach described, we conducted a mock clinical validation using the same developmental datasets employed for algorithm training (the final validation process would be performed using a different dataset that is distinct from the development data). A virtual cohort of 629 total datapoints was constructed using data from various radiation accidents, publicly available data from unirradiated subjects, and simulated data for irradiated conditions. As described above, this example cohort was structured to align with the Buddemeier NYC dose distribution model (using the MDZ + LDZ + DFZ distribution). We used this virtual cohort to assess the CellRADx algorithm’s performance, using a random sampling of available timepoints, one from each subject, to avoid use of repeated measures. This random sampling approach was repeated 5 times and the results were averaged. The results of this process are summarized in Table 1, which reports the number of RWD, modeled and total data points, the accuracy, and the mean number of correct results, all within each individual dose bin. The sensitivity of 92% is calculated by dividing the total number of correct calls in all dose bins >2 Gy (60) by the total count in all dose bins >2 Gy (65). Similarly, the specificity of 78% is calculated by dividing the total number of correct calls in all dose bins between 0 and 2 Gy (440) by the total count in all dose bins between 0 and 2 Gy (564).

Table 1.

Example Virtual Cohort Composition and CellRADx Performance Assessment Results

Dose bin (Gy)	Real-world datapoints	Modeled datapoints	Total datapoints	Mean correct results	Accuracy
0	389	0	389	352.2	90.5%
>0.0-0.5	9	69	78	51.4	65.9%
>0.5-1.0	8	54	62	24.8	40.0%
>1.0-1.5	6	15	21	7.4	35.2%
>1.5-2.0	5	9	14	4.0	28.6%
>2.0-2.5	3	7	10	7.8	78.0%
>2.5-3.0	7	0	7	6.0	85.7%
>3.0-4.0	4	7	11	10.0	90.9%
>4.0-5.0	6	3	9	8.6	95.6%
>5.0-6.0	5	2	7	7.0	100.0%
>6.0-7.0	0	6	6	6.0	100.0%
>7.0-8.0	1	4	5	4.8	96.0%
>8.0-9.0	0	5	5	5.0	100.0%
>9.0-10.0	2	3	5	4.8	96.0%
Total	445	184	629	Specificity:	78.0% (95% CI: 74.3-81.3%)
Proportion	71%	29%	100%	Sensitivity:	92.3% (95% CI: 83.0-97.5%)

A virtual cohort of 629 total datapoints was constructed using data from various radiation accidents, publicly available data from unirradiated subjects, and simulated data for irradiated conditions. The virtual cohort was used to assess the CellRADx algorithm’s performance, using a random sampling of available timepoints, one from each subject, repeated 5 times. The table reports the number of real world, modeled and total data points, the algorithm’s accuracy, and the mean number of correct results, all within each individual dose bin, as well as the overall clinical sensitivity and specificity.

To assess performance with various subpopulations that may confound the test, we followed a similar approach. Each of these populations was previously identified as having reference ranges for select CBC measurands that differ from apparently healthy adults, as well as having meaningful prevalence within the intended use population. However, as there was no radiation accident data from these populations, in silico data was used exclusively to represent the irradiated conditions. This data was generated using the process described above, wherein data from unirradiated representative subjects acquired from NHANES was processed through the cell depletion model, using unaltered model parameters, to create CBC values at various dose levels and times post-exposure. The same randomized sampling and analysis methods were then used to assess algorithm performance with these populations. In nearly all cases, the sensitivity target was achieved (range of 89%-98%). However, in a number of cases, the specificity was lower than our target (range of 69%-90%), indicating that additional CellRADx algorithm refinement is likely necessary. We anticipate that changes can be made to the threshold used for specific populations to improve performance, as well as more extensive changes to the underlying parameters in the algorithm. Results for each subpopulation are provided in Table S2.

DISCUSSION

The CellRADx software device is specifically designed to address a large-scale nuclear incident, wherein hundreds of thousands to millions of people may seek medical care for possible radiation exposure and overwhelm the local healthcare system. In this type of scenario, medical response personnel will need an easy way to effectively triage the population and identify those who truly require immediate medical attention.

Our sensitivity and specificity performance targets (90% and 80%, respectively) are based on considerations of the benefits and risks associated with the use of the device. Individuals identified as positive are likely to undergo further testing, including quantitative biodosimetry methods, to confirm their exposure level and further inform appropriate medical treatments. They may also receive radiation medical countermeasures (MCMs), such as Neulasta, Neupogen, Leukine (or one of the FDA-approved biosimilars for these growth factors), or Nplate, and other treatments (e.g., antibiotics, fluids). Conversely, those individuals identified as negative will likely be removed from the emergency medical care queue, as their care can be provided by non-emergency medical resources and reduce the burden on emergency resources. However, false negative results could lead to an individual not receiving early treatment for survivable radiation sickness (assuming MCMs would have been administered) and thus the individual may suffer more serious health impacts. Given that the side effects of radiation MCMs for H-ARS given to unexposed or low dose patients are typically mild and occur in a minority of subjects,^22-25 the benefits of early identification that can inform administration of H-ARS MCMs and other medical care to individuals with significant exposure far outweigh the risks associated with false identification.

This consideration warrants performance objectives that are intentionally biased towards higher sensitivity to reduce numbers of false-negative results. This approach aligns with other FDA-cleared screening tests, such as the Banyan BTI test (DEN170045) and the Abbot i-STAT test (K201778) for traumatic brain injury, which achieve high sensitivity (> 95%) at the expense of low specificity (< 41%). However, the approach is only suitable when resources such as healthcare personnel, facilities, and MCMs are readily available to accommodate the potential high rate of false positive results. A mass-casualty radiological or nuclear disaster is anticipated to involve severely resource-constrained environments, wherein specificity becomes more meaningful to reduce the application of medical resources to those persons misclassified as positive. Therefore, the challenge in a large-scale nuclear incident is balancing the competing objectives of maximizing the sensitivity although keeping the specificity at levels appropriate for the available resources. In consideration of these objectives, performance targets of 80% specificity and 90% sensitivity were established.

As expected, diagnostic accuracy is lowest for CellRADx at exposure levels near the threshold. As with any qualitative test, there is nearly always a point on the probability of detection curve where the accuracy is 50%. Typically, this 50% point aligns with the threshold used to make positive vs negative calls. However, in the case of CellRADx, that 50% accuracy point is lower than the 2 Gy threshold (a bit less than 1 Gy) because the algorithm is intentionally biased to favor sensitivity, so performance is lowest at doses just below 2 Gy. As noted above, this will lead to a higher rate of false positives in people who had a mild level of exposure, but it is an acceptable tradeoff because the sensitivity performance is excellent, ensuring the vast majority of subjects who received a significant dose are correctly identified as positive.

There are several opportunities for future research. For example, much of the radiation accident data, on which the cell depletion model is based, is many decades old with largely unquantified uncertainties in estimated dose values and CBC measurements. Engagement with radiobiologists, medical physicists, and other experts would increase scientific consensus around the cell depletion model assumptions and help quantify uncertainties. Additionally, participation in exercises to gain user feedback under simulated conditions would improve the application and the likelihood of future adoption. Further, we may consider incorporation of additional biomarkers to increase accuracy, but this will further increase regulatory complexity, and future efforts must include dedicated engagement with regulatory agencies to agree on appropriate methods for clinical validation.

CONCLUSION

In summary, we have developed a novel approach to clinically validate the CellRADx software and its underlying analysis algorithm. Using CBC data collected from various unirradiated populations and published computational models of blood cell count changes over time after radiation exposure, we can generate simulated CBC data at any time post-exposure and at any dose level within our defined range. This process can be performed for virtually any subpopulation of interest using CBC data from representative subjects. Applying this collection of in silico data to the expected distribution of radiation exposure levels based on published nuclear blast modeling effectively simulates enrollment of a clinical population after a nuclear detonation. Processing this collection of data through the CellRADx software emulates a conventional, prospective clinical trial. The performance of the device can thus be validated to support FDA authorization by demonstrating that the device works as intended and is safe and effective.

SUPPLEMENTARY MATERIAL

Supplementary material is available at Military Medicine online.

FUNDING

This project has been funded in whole or in part with Federal funds from the National Institute of Allergy and Infectious Diseases, under award U01AI148316, and builds on previous work funded by the Biomedical Advanced Research and Development Authority, under contract 75A50122C0050.

Supplement Sponsorship

This article appears as part of the supplement “Proceedings of the 2025 Military Health System Research Symposium,” sponsored by the Assistant Secretary of Defense for Health Affairs.

CONFLICT OF INTEREST STATEMENT

The authors do not have a conflict of interest.

DATA AVAILABILITY

The derived data generated in this research will be shared upon reasonable request to the corresponding author.

INSTITUTIONAL REVIEW BOARD

Not applicable. The collection of deidentified data is not considered human subjects research and is exempt from IRB review (confirmed with an IRB).

INSTITUTIONAL ANIMAL CARE AND USE COMMITTEE

Not applicable.

INSTITUTIONAL CLEARANCE

AFRRI/USUHS granted approval for public release on 8 Oct 2025.

INDIVIDUAL AUTHOR CONTRIBUTION STATEMENT

All authors contributed to the design and implementation of the research, to the analysis of the results, and to the review of the manuscript. K.S.D. provided oversight from the manuscript’s inception through the final review. All authors read and approved the final manuscript.