OBJECTIVES:

Diagnosis rates for hepatitis C virus (HCV) remain low, despite advances in treatments and recommendations for widespread screening. This study assesses the potential of artificial intelligence (AI) to provide a targeted population for screening.

METHODS:

This is a retrospective observational study based on US ambulatory electronic medical records covering 70M patients across primary and specialty care. The study period was January 2015 to August 2020. Patients with an initial diagnosis of HCV during a 12 month observation period were labelled as ‘HCV’, whilst patients with no HCV diagnosis throughout the entire study period were labelled as ‘non-HCV’. The index date was the beginning of the observation period and the lookback date was 24 months earlier. The timing and frequency of pre-index events were captured for HCV-relevant medical history. An AI algorithm, gradient boosted trees, was used to predict HCV vs non-HCV. Positive Predictive Values (PPVs) and associated sensitivities were computed based on a hold-out sample. Universal screening was used as the baseline for benchmarking the performance of the AI algorithm alongside more targeted approaches: screening the 1946-1964 birth cohort and people who inject drugs (PWID).

RESULTS:

26,731 HCV and 14.3M non-HCV patients met the selection criteria. At 5% sensitivity, the AI algorithm achieved a PPV 43 times higher than universal screening. At sensitivity levels matched to targeted screening initiatives, PPVs were:

• 7 and 19.9 times higher than universal screening for PWID screening and the AI algorithm, respectively.
• 4 and 4.4. times higher than universal screening for the birth cohort and the AI algorithm respectively.

The key features identified by the algorithm were age, gender, pain medications, cholesterol, and antidepressants.

CONCLUSIONS:

This study provides encouraging evidence that routinely collected medical data coupled with AI provides a route to personalised targeting for screening that could complement existing screening initiatives.