An explainable, bias-audited machine learning model to detect hidden hepatitis C from region-wide real-world health records

An explainable Boosting Machine trained on Andalusian health records detects undiagnosed hepatitis C with AUC 0.837 in an independent 2024 population sample.

• Why it matters: Because 25–30% of active HCV infections in Spain remain undiagnosed, scalable screening needs models that use routine data while controlling bias and maintaining interpretability.
• What they did: A retrospective, population-based cohort in the Andalusian Population Health Database matched 25,028 diagnosed cases (2017–2024) to controls up to 10:1, trained on 2017–2022 and thresholded in 2023 to guarantee ≥60% sensitivity.
• The result: In the independent 2024 test set, sensitivity was 63.5% and specificity 91.8% (PPV 7.0%, NPV 99.6%), with key predictors including minimum alanine aminotransferase, HIV infection, non-tobacco substance dependence, and platelet-count statistics.

AI summary of the abstract · may contain errors · not medical advice · read the paper.