Dual-energy X-ray absorptiometry (DXA) is a primary modality for assessing bone mineral density (BMD). The authors developed an agentic research workflow, DXA Agent, to coordinate planning, feature refinement, tools, provenance tracking and hypothesis-generating interpretation for DXA outcomes. The workflow explicitly considers measurement burden (cost and complexity of acquiring features) alongside predictive performance and reproducibility.
DXA Agent is presented as an agentic modelling pipeline that integrates automated and human-directed steps: planning of modelling tasks, iterative feature-model refinement, tool integration, provenance recording and interpretation to generate hypotheses. The workflow was designed to be inspectable and to enable evaluation of trade-offs between predictive accuracy and the burden of measurements required to generate predictions.
Models were developed and tested independently in two large cohort resources:
The authors report results across 20 bone mineral density sites in UK Biobank and 3 sites in NHANES.
The study compared two broad strategies: a cost-efficient strategy that limits measurement burden and a no-limit strategy that permits use of all available features. Models developed under these strategies were compared to the best conventional comparator models for each target site. The workflow included steps for feature-model refinement and recorded provenance to support inspectability and reproducibility.
Continuous predictions were evaluated using root mean squared error (RMSE) and R2. Classification tasks (for osteoporosis) were assessed using AUROC and precision-recall area under the curve (PR-AUC). Comparisons were made to conventional models, described by the authors as the best available comparator for each outcome and site.
Across the 20 UK Biobank and 3 NHANES bone mineral density sites, the cost-efficient models achieved lower RMSE and higher R2 than the best conventional comparator models. The authors report median relative RMSE reductions of 10.9% and 9.9%, respectively, across the evaluated sites. These reductions reflect improved continuous prediction accuracy for BMD when using the measurement-burden-aware modelling approach in their retrospective, cohort-internal experiments.
Classification results were task-dependent. For UK Biobank osteoporosis classification, the average discrimination metrics reported were AUROC 0.839 and PR-AUC 0.182. For NHANES classification tasks, performance was reported to be comparable with conventional models rather than uniformly superior. The authors note variation in classification outcomes by cohort and task.
The study assessed the incremental value of including higher-burden (more costly or complex to obtain) features. The authors report that inclusion of higher-burden features did not consistently improve predictive performance across sites. This finding underpins the value proposition of the cost-efficient modelling strategy embedded in DXA Agent.
DXA Agent emphasizes inspectability: models, feature selection decisions and tool use are tracked to permit interpretation and hypothesis generation post hoc. The workflow is intended to make modelling choices and their provenance transparent, supporting reproducibility and enabling investigators to weigh measurement burden against gains in predictive performance.
These analyses are retrospective and cohort-internal; the authors explicitly state that independent prospective validation is required. Ethics and data access statements provided by the authors include UK Biobank ethical approval by the North West Multi-centre Research Ethics Committee (reference 21/NW/0157) and UK Biobank data access under Application 45925. NHANES data availability is referenced via the public NCHS resource. The authors declared no competing interests.
In these retrospective analyses across two population cohorts, the measurement-burden-aware DXA Agent workflow produced cost-efficient models that reduced RMSE and increased R2 relative to conventional comparators across multiple DXA sites, with reported median relative RMSE reductions of 10.9% (UK Biobank) and 9.9% (NHANES). Classification performance varied by task and cohort; osteoporosis discrimination in UK Biobank averaged AUROC 0.839 and PR-AUC 0.182. Higher-burden features did not reliably improve prediction. The authors position DXA Agent as an inspectable research workflow that requires independent prospective validation before broader adoption.
This preprint was posted September 01, 2026. The copyright holder granted medRxiv a license to display the preprint in perpetuity under a CC-BY 4.0 International license.