S3013 From Black-Box AI to Transparent Decision Support: Evaluating DeepSeek’s Reasoning for Enhancing AI Adoption in Clinical Care

Abstract

Introduction: The integration of artificial intelligence (AI) into gastroenterology and hepatology is limited by the opacity of traditional “black-box” large language models (LLMs). DeepSeek, an large language models with an intrinsic Chain of Thought framework, autonomously generates transparent, stepwise reasoning, addressing trust and usability barriers. This study evaluates DeepSeek’s performance in high-acuity gastrointestinal/hepatological cases. Methods: Four expert-reviewed cases (hepatorenal syndrome, esophageal variceal bleeding with hepatic encephalopathy, Boerhaave syndrome with septic shock, and acute pancreatitis with limb ischemia) were assessed by 6 physicians (gastroenterology attendings, fellows, and internal medicine residents). Responses were scored across 5 domains: reasoning clarity, guideline alignment, prioritization of critical factors, management consistency, and clinician confidence. A structured survey quantified transparency, usability, and workload impact. Statistical analysis included interrater reliability (Cohen’s κ) and ANOVA. Results: DeepSeek achieved 85% reasoning clarity and 81.7% guideline alignment, with 93.3% trust among attendings versus 83.3% for residents. It prioritized life-saving interventions effectively (82.5%) and maintained reasoning-management consistency (78.3%). Evaluators preferred DeepSeek over black-box models (86.7%) even with higher accuracy, reporting 73% reduced cognitive workload and 86.7% time savings. Junior clinicians benefited from structured diagnostic pathways, while attendings used it for decision validation. Strong interrater reliability (κ = 0.82) and experience-based rating differences (P < 0.01) highlighted adaptability across training levels. Conclusion: DeepSeek’s Chain of Thought framework enhances clinical AI adoption through transparency and workflow efficiency. Explainability is as critical as accuracy in high-stakes specialties. Future integration of retrieval-augmented generation for real-time guideline access and domain-specific reinforcement learning could refine context-aware decision support while adhering to data privacy standards. This study underscores the necessity of balancing technical performance with clinician-centric design to advance AI’s role in gastroenterology and hepatology.

Ähnliche Arbeiten