An explainable, bias-audited machine learning model to detect hidden hepatitis C from region-wide real-world health records
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Although direct-acting antivirals have transformed hepatitis C virus (HCV) treatment, 25-30% of active infections in Spain remain undiagnosed. We aimed to develop and validate an interpretable machine-learning model to identify undiagnosed HCV infection using only data already recorded within the health system.
Materials and Methods
We conducted a retrospective, population-based cohort study using the Andalusian Population Health Database (BPS). Individuals diagnosed with HCV between 2017 and 2024 (n=25,028) were matched, up to 10:1, to controls by birth year, sex, and health district. An Explainable Boosting Machine was trained on 2017-2022 data, with a threshold set in a 2023 validation set to guarantee ≥60% sensitivity, then evaluated unmodified in an independent, non-matched, population-representative 2024 sample approximating real-world screening. Predictors comprised chronic-disease diagnoses and laboratory summary statistics recorded ≥30 days before the index date; age and sex were excluded as matching variables. Predictors reflecting socioeconomic proxies or matching artefacts (pharmaceutical co-payment category, a zone-level deprivation index) were removed after explainability review.
Results
In the independent test set, the model achieved an AUC of 0.837 (0.826 in validation), sensitivity 63.5%, specificity 91.8%, positive predictive value 7.0%, and negative predictive value 99.6%. Leading predictors were minimum alanine aminotransferase, HIV infection, dependence on non-tobacco substances, and platelet-count statistics, forming a clinically coherent, laboratory- and comorbidity-driven profile.
Conclusion
An explainable model built exclusively from routine electronic health record data can identify undiagnosed HCV infection with discrimination comparable to previously published tools, evaluated under conditions approximating true population prevalence. Bias-auditing was feasible without sacrificing interpretability, supporting a scalable, equity-conscious instrument for the final phase of HCV elimination.
