OBJECTIVES: To explore and evaluate various reasoning algorithms of large language models (LLM) and their application in HEOR.

METHODS: This study reviews and conducts a theoretical examination of existing advanced AI reasoning algorithms, assessing their potential for use in HEOR by enabling more complex, iterative, and nuanced decision-making processes. The evaluation focuses on how these algorithms can facilitate more dynamic and adaptive reasoning in HEOR. Four reasoning algorithms were assessed: Basic Input-Output (IO), Chain of Thoughts (CoT), Multiple Chains of Thoughts (CoT-SC), Tree of Thoughts (ToT), and Graph of Thoughts (GoT).

RESULTS: Each reasoning model offers unique advantages: The Basic IO approach, which is commonly used, provides straightforward answers, however, is linked with limited reasoning abilities and often leads to hallucinations. The Chain of Thoughts introduces intermediate steps in reasoning, allowing for more transparent and traceable AI reasoning or decisions. The Multiple Chain of Thoughts enhances this by evaluating multiple reasoning paths, selecting the most robust outcome. The Tree of Thoughts expands further, generating and backtracking through multiple thoughts to explore various decision avenues extensively. Finally, the Graph of Thoughts uses graph-based transformations to aggregate and refine thoughts, offering a sophisticated tool for complex problem-solving in HEOR. An integrated human-in-loop approach, alongside these reasoning algorithms was employed.

CONCLUSIONS: Advanced AI reasoning algorithms, beyond IO approach, while keeping human in loop has potential to provide a deeper analysis and more refined reasoning outcomes, limiting hallucinations and building transparency in HEOR studies. These models facilitate a shift from linear to more complex, networked decision-making processes, greatly enhancing the analytical capabilities. Implementing these algorithms, will require careful consideration of computational resources and capacity building to fully harness their potential, ensuring that AI reasoning evolves into a critical component in HEOR applications.