Reconstructing strontium-90 intake in beagles using neural networks: a data-driven assessment of historical inhalation records

Dose estimation in response to internal radionuclide exposures requires reconstruction of the initial intake activity, which is frequently unknown due to the absence of<i>a priori</i>data. In such scenarios, intake is inferred from bioassay measurements obtained at one or more time points post-exposure. Reconstructing an initial intake from bioassay relies on biokinetic models that describe the body distribution and clearance of the toxicant. These models typically employ first-order differential equations with generalised population parameters, which do not capture individual variation in metabolism or anatomy. Thus, reconstruction of initial intakes is affected by multiple sources of stochasticity, including physical deposition of the inhaled radionuclide, detection system uncertainty, and inter-individual physiological variability. The capacity of machine learning (ML) algorithms to model highly non-linear and often stochastic processes makes them appropriate for augmenting intake reconstruction. This study applies artificial neural networks to estimate the initial intake activity of⁹⁰Sr inhaled by beagles. Model performance and sensitivity to input data quality were assessed through inclusion of individual-specific features, such as age, weight, and sex. Three data regimens were systematically designed, each with distinct pre-processing pipelines and model complexity. The first regimen demonstrates feasibility of intake reconstruction using bioassay measurements taken within 14 days post-exposure. The second regimen demonstrates that summary statistics of retention functions in historical records lack sufficient resolution for individualised ML modelling. The third regimen shows that historical dose estimates, despite limitations in resolution and methodology, can be used as surrogate features when multiple post-exposure time points are available. Root mean squared error was used to evaluate prediction error, while a custom metric, the variance relative difference, quantified model bias. In addition to evaluating predictive performance, this study assesses the integrity and usability of historical data from⁹⁰Sr beagle inhalation experiments conducted at the Inhalation Toxicology Research Institute between 1966 and 1987.