Predicting explainable dementia types with LLM-aided feature engineering.

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Machine Learning for Healthcare

Background:

• Healthcare generates vast clinical data, necessitating advanced analytics.
• Current AI models, including Large Language Models (LLMs), struggle with explainability and reliability in clinical settings.
• High-stakes medical applications demand interpretable and trustworthy AI solutions.

Purpose of the Study:

• To develop an LLM-aided feature engineering method for enhanced interpretability in healthcare AI.
• To extract clinically relevant features from medical texts.
• To improve the performance and efficiency of AI models in medical data analysis.

Main Methods:

• Utilized Large Language Models (LLMs) for feature engineering from the Oxford Textbook of Medicine.
• Converted clinical notes into concept vector representations.
• Employed a linear classifier for analysis and compared against n-gram Logistic Regression and GPT-4 baselines.
• Investigated Text Embeddings for computational efficiency.

Main Results:

• Achieved 0.72 accuracy using the LLM-aided approach, surpassing traditional n-gram Logistic Regression (0.64) and GPT-4 (0.48) baselines.
• Focused on extracting high-level clinical features for improved interpretability.
• Reduced overall time and cost by 97% through the use of Text Embeddings.

Conclusions:

• The proposed LLM-aided feature engineering enhances AI interpretability and accuracy in healthcare.
• This method offers a cost-effective and efficient solution for clinical data analysis.
• The approach demonstrates the potential of LLMs in overcoming explainability challenges in medical AI.