Lightweight Language Models are Prone to Reasoning Errors for Complex Computational Phenotyping Tasks.

Area of Science:

• Biomedical Informatics
• Artificial Intelligence

Background:

• Computational phenotyping is essential for cohort identification but is time-intensive due to manual data review.
• Previous studies showed limitations of LLMs in complex phenotyping tasks, particularly with multiple therapies.

Purpose of the Study:

• To evaluate the reasoning capabilities of lightweight LLMs in computational phenotyping.
• To enhance the PHEONA framework for assessing faulty reasoning in LLMs.

Main Methods:

• Assessed three lightweight LLMs (DeepSeek-r1, Mistral Small, Phi-4) for phenotyping accuracy.
• Utilized prompt modifications to identify explanation correctness and unfaithfulness errors.
• Expanded the PHEONA framework to include faulty reasoning evaluation.

Main Results:

• Reasoning errors, including explanation correctness and unfaithfulness, were prevalent across all tested LLMs.
• DeepSeek demonstrated the smallest accuracy impact after prompt modifications compared to Mistral and Phi.
• The enhanced PHEONA framework successfully identified pervasive reasoning errors.

Conclusions:

• Reasoning errors are ubiquitous in LLM responses for complex tasks like computational phenotyping.
• The enhanced PHEONA framework is vital for LLM evaluation, highlighting the need for improved interpretability methods.