AI-Driven Prediction of Chest CT Radiation Doses: Establishing BMI-Based Diagnostic Reference Levels and Patient–Factor Correlations for Machine-Learning Models

Zuhal Y. Hamd, Mohamed Abuzaid, Mohamed Alharbi, Nissren Tamam, Amal I. Alorainy, Lena Alrujaee, Najla Almutairi, Aljouharah Abdullah Alyagoub

Background and aim: Chest CT is a major contributor to population radiation exposure. Conventional, pooled diagnostic reference levels (DRLs) do not account for inter-individual variability in body habitus and are typically used retrospectively. We evaluated dose behavior in adult chest CT, derived BMI-stratified local DRLs, and developed models to enable AI-assisted, prescan dose prediction. Methods: Consecutive adult chest CT examinations from a single center were analyzed. Dose indices (CTDIvol, DLP) and patient factors (BMI, weight, height, age, sex; scan length and planned technical parameters where available) were extracted. DRLs were defined as the 75th percentile overall and within BMI categories (underweight, normal, overweight, and obese). Group differences were assessed using non-parametric tests; associations were examined using correlation analysis. Supervised learning (e.g., Random Forest, Gradient Boosting) was trained to predict CTDIvol and DLP from routinely available variables. Results: BMI-stratified DRLs increased monotonically with habitus: underweight 444.95 mGy·cm/9.60 mGy; normal 513.00/11.55; overweight 756.08/14.65; obese 931.60/20.25 (DLP/CTDIvol). Differences across BMI groups were significant for DLP (H = 31.53, p < 0.001) and CTDIvol (H = 33.61, p < 0.001). DLP correlated moderately with weight and BMI (r ≈ 0.54–0.56, p < 0.001), with a weaker association for age; height was not a meaningful predictor. No sex-based differences in CTDIvol or DLP were observed. Predictive models estimated CTDIvol and DLP with high performance (R2 up to ~0.79 and ~0.77, respectively), enabling comparison of predicted dose against BMI-matched DRLs before acquisition. Conclusions: Size-aware, BMI-stratified DRLs provide clinically interpretable investigation levels that avoid pitfalls of pooled benchmarks. Coupled with robust prediction of individualized dose from routine variables, this framework supports a shift from retrospective audit to prospective, point-of-care dose governance and protocol optimization in chest CT.