Integrative and interpretable machine learning framework for early non-invasive detection of clinically significant liver fibrosis.

Background: Liver fibrosis is a critical predictor of adverse liver-related outcomes, yet its early detection remains challenging. This research aimed to construct a predictive model for the early detection of clinically significant liver fibrosis, and identify its key risk factors, using clinical data from the NHANES 2017-2020 cycles. Methods: We analyzed a cohort of 6,164 participants, classifying them into a clinically significant liver fibrosis group and control group based on liver stiffness measurements. Data were preprocessed, with imputation for missing values and normalization of numerical variables. We employed recursive feature elimination and benchmarked 29 machine learning models, ultimately selecting the Gamboost model due to its excellent performance. The model's robustness was validated with an external dataset of 684 samples. Results: Key predictors included body mass index (BMI), aspartate aminotransferase (AST), and waist circumference (WC), among others. In the training cohort, the model reached an AUC of 0.824 (95% CI: 0.798-0.849), while in the testing cohort it yielded 0.872 (95% CI: 0.838-0.905), maintaining robust performance with an AUC of 0.848 (95% CI: 0.798-0.898) during external validation. Furthermore, SHAP and PDP analyses revealed that elevated AST and WC synergistically increased the predicted risk of fibrosis. Conclusion: By utilizing routinely accessible clinical data, our interpretable machine learning model offers a highly practical tool for the early identification of clinically significant liver fibrosis. This non-invasive approach can be readily implemented in primary care, facilitating timely interventions and supporting precision medicine initiatives.