Unveiling patterns in clinical data: exploring the role of large language models and clustering algorithms

Area of Science:

• Clinical informatics
• Artificial intelligence in healthcare
• Data science

Background:

• Large Language Models (LLMs) excel in natural language processing but are underutilized for structured clinical data.
• Analyzing clinical data requires methods that preserve structural integrity and enhance predictive capabilities, especially in resource-limited settings.

Purpose of the Study:

• To investigate if LLM-generated embeddings can maintain the structure of clinical datasets.
• To assess the enhancement of predictive modeling using LLM embeddings in resource-constrained environments.

Main Methods:

• Dimensionality reduction (PCA, t-SNE) and clustering (k-means) were used to compare original data structures with LLM embeddings.
• Evaluation involved cosine similarity, AUC, and R-squared on synthetic and real-world clinical datasets (UCI medical, endocarditis).
• Multiple LLM architectures (BERT, RoBERTa, Llama 2, E5-small) were assessed for predictive accuracy and computational efficiency.

Main Results:

• LLM embeddings closely replicated original data structures, with high cosine similarity (e.g., BERT at 0.95 for linear data).
• Predictive performance improved with increased subject variable ratio (SVR), identifying distinct performance groups.
• Different LLM architectures showed varying trade-offs between computational cost and predictive performance.

Conclusions:

• LLMs can enhance structured data analysis and identify optimal conditions (e.g., SVR thresholds) for practical application.
• Context-specific LLM selection is crucial due to performance and computational cost variations.
• LLMs enable repurposing clinical data for individualized questions, advancing precision medicine and data-driven decisions.