Developing and Externally Validating Explainable Machine Learning for Elastography-Defined Steatotic Liver Disease

Explainable machine learning is widely applied to non-invasive detection of steatotic liver disease, but published pipelines are rarely tested against a properly specified regression comparator, rarely validated independently, and almost never examined for measurement artifact in the outcome label itself. This protocol addresses those gaps in a form that can be re-executed end-to-end on public data. The procedure uses the National Health and Nutrition Examination Survey (NHANES) 2017 to March 2020 release for development and the August 2021 to August 2023 release as a frozen external validation set. Steps include verification of variable codes against the raw transport-file headers, cohort construction, recoding of skip-pattern missingness prior to multiple imputation, nested cross-validation with tuning confined to training folds, design-weighted performance estimation, paired-bootstrap comparison, SHapley Additive exPlanations (SHAP) analysis with stability checks, and ten pre-specified sensitivity analyses. Representative results show that routine clinical data discriminate VCTE-defined steatosis at an area under the receiver operating characteristic curve (AUROC) of 0.843 and advanced fibrosis at 0.837, both of which were reproduced in external validation, and that these values greatly exceed those of FIB-4, which discriminated at 0.565. Gradient boosting did not outperform penalized logistic regression for steatosis and showed at most a modest advantage for fibrosis. A probe-selection test built into the protocol revealed that the XL elastography probe reads the controlled attenuation parameter (CAP) 22.9 dB/m higher than the M probe at equivalent body mass index, and that removing this coupling reduced steatosis discrimination from 0.843 to 0.810. The protocol applies to any NHANES prediction question and, with the weighting steps adapted, to other complex-survey datasets. Its distinguishing features are the artifact test, the cross-method explainability check, and full pre-specification of sensitivity analyses.