Decision Tree Extraction for Clinical Decision Support System With If-Else Pseudocode and PlanSelect Strategy.

Area of Science:

• Artificial Intelligence
• Medical Informatics

Background:

• Decision trees are vital for clinical decision support systems, but current extraction methods using large language models struggle with structural understanding and node content accuracy.
• This leads to noisy extracted trees, hindering their effectiveness in clinical decision-making.

Purpose of the Study:

• To propose a novel framework for accurate and complete decision tree extraction from medical knowledge.
• To improve the performance of clinical decision support systems by enhancing decision tree retrieval.

Main Methods:

• A two-stage framework was developed: first, using If-Else pseudocode with specific constraints to guide large language model (LLM) output for decision tree structure.
• Second, a PlanSelect node-filling strategy was introduced to match extracted triplets with pseudocode sub-sentences, incorporating observation, plan, action, and answer reasoning steps.

Main Results:

• The proposed method demonstrated superior performance over state-of-the-art approaches on both public and newly constructed datasets (EMDT), achieving lower ER metrics.
• Specifically, improvements of 1.37% and 1.54% were recorded on the Text2DT and EMDT datasets, respectively.

Conclusions:

• The novel framework effectively addresses challenges in decision tree extraction, producing more accurate and complete representations of medical knowledge.
• Extracted medical decision trees significantly improved performance on clinical decision-making tasks like CMB-Clin and MedQA.