NPJ Digit Med. 2026 May 9. doi: 10.1038/s41746-026-02718-y. Online ahead of print.
ABSTRACT
Artificial intelligence (AI)-based screening tools show promise for early identification of chronic liver disease (CLD), yet their effectiveness in real-world settings may depend on clinician response to AI-generated recommendations. We performed a post hoc analysis of the intervention arm of the pragmatic, cluster-randomized DULCE trial, in which primary care clinicians received electrocardiogram-based machine learning (ECG-ML) alerts indicating elevated risk for CLD. Clinicians were categorized as high engagement (HE; top quartile) or low engagement (LE), and diagnostic yield was defined as the proportion of ECG-ML-positive cases with confirmed CLD. Among 110 clinicians receiving ≥1 alert (1385 ECG-ML-positive patients), overall engagement was 29.8%. HE was associated with higher detection of advanced CLD (OR 2.12, 95% CI 1.36-3.30; p = 0.001) and any CLD (OR 2.59, 95% CI 1.83-3.68; p < 0.001) compared with LE. Diagnostic yield was 10.6% versus 2.9% for advanced CLD and 22.3% versus 5.0% for any CLD in HE versus LE (OR 2.99, 95% CI 1.73-5.16; p < 0.001 and OR 3.74, 95% CI 2.44-5.75; p < 0.001, respectively). These findings suggest that the effectiveness of AI-based screening may depend not only on algorithm performance but also on clinician engagement with AI recommendations and highlight the importance of accounting for engagement when designing and interpreting AI-enabled clinical trials. ClinicalTrials.gov NCT05782283.
PMID:42106573 | DOI:10.1038/s41746-026-02718-y