AI-assisted vs. textbook-based vs. blended learning for acute abdomen diagnosis: a retrospective cohort study of emergency interns

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Background Large language models (LLMs) are reshaping medical education. Objective To examine whether learning approach—AI-assisted, textbook-based, or blended—was associated with diagnostic accuracy for acute abdominal conditions among emergency interns. Methods We reviewed 720 clinical decisions by 72 emergency interns at a tertiary center over 12 months. Interns were classified into three groups: textbook-based ( n = 27), AI-assisted ( n = 21), and blended ( n = 24). Propensity score matching with pair-stratified GEE addressed selection bias and clustering. Results Diagnostic accuracy was 73.0% (textbook), 82.4% (AI-assisted), and 87.1% (blended) ( p < 0.001). Blended learning showed the largest advantage over textbook (adjusted OR = 3.21, 95% CI: 1.87–5.51, p < 0.001), followed by AI-assisted (adjusted OR = 1.68, 95% CI: 0.99–2.85, p = 0.053). Matched analyses confirmed the benefit for blended learning ( p = 0.011). Subgroup analyses suggested larger effects for severe cases (OR = 4.92) and atypical presentations (OR = 3.85), with an E-value of 5.88 supporting robustness against unmeasured confounding. Conclusion Blended learning combining AI tools with traditional resources showed the strongest association with diagnostic accuracy, suggesting that integrated rather than exclusive approaches may better support clinical reasoning in emergency training.