Artificial intelligence for clinical reasoning: the reliability challenge and path to evidence-based practice.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Decision Support

Background:

• Generative artificial intelligence (AI), especially large language models (LLMs), offers potential to transform clinical reasoning.
• Current LLMs demonstrate capabilities like passing medical exams but face a "reliability challenge" in replicating human decision-making.
• Limitations include mimicking reasoning rather than true logic and using outdated or non-regional data, impacting clinical relevance.

Purpose of the Study:

• To address the reliability challenge of AI in clinical reasoning.
• To propose a synergistic model integrating physician expertise with advanced AI.
• To outline requirements for trustworthy AI in medical practice.

Main Methods:

• Conceptual analysis of AI capabilities and limitations in clinical settings.
• Advocacy for a paradigm shift towards transparent and interpretable AI.
• Identification of necessary AI system attributes: real-time data integration, context-specificity, and explainable architectures.

Main Results:

• LLMs can achieve high diagnostic accuracy but lack genuine logical reasoning.
• Current AI models often rely on data that is not current or locally relevant.
• Explainable AI architectures, such as multi-step reasoning or knowledge graphs, are crucial for transparency.

Conclusions:

• Reliable AI for clinical reasoning necessitates a blend of technological advancement and human oversight.
• AI systems must integrate real-time, context-specific evidence and align with local healthcare constraints.
• Ethical principles of beneficence and non-maleficence must guide AI development for evidence-based, patient-centered care.