Contextual Phenotyping of Pediatric Sepsis Cohort Using Large Language Models.

Area of Science:

• Computational biology
• Medical informatics
• Artificial intelligence in healthcare

Background:

• Personalized care and resource optimization require effective patient subgroup clustering.
• Traditional methods face challenges with high-dimensional, heterogeneous healthcare data and lack contextual understanding.
• Pediatric sepsis datasets from low-income countries present unique challenges for analysis.

Purpose of the Study:

• To evaluate Large Language Model (LLM)-based clustering against classical methods for pediatric sepsis patient subgroups.
• To assess the performance of different LLM embedding models and clustering objectives.
• To determine the potential of LLM clustering for contextual phenotyping in resource-limited settings.

Main Methods:

• Patient records were serialized into text and clustered using LLM embeddings (LLAMA 3.1 8B, DeepSeek-R1-Distill-Llama-8B, Stella-En-400M-V5) with K-means.
• Classical methods (K-Medoids) were applied to dimensionality-reduced mixed data (UMAP, FAMD).
• Clustering quality was assessed using Silhouette scores and statistical tests on a pediatric sepsis dataset (2,686 records).

Main Results:

• Stella-En-400M-V5 achieved the highest Silhouette Score (0.86).
• LLAMA 3.1 8B with a clustering objective identified distinct patient subgroups based on nutritional, clinical, and socioeconomic profiles.
• LLM-based methods demonstrated superior performance over classical techniques by capturing richer context.

Conclusions:

• LLM-based clustering effectively captures contextual information and outperforms classical methods for heterogeneous healthcare data.
• LLM clustering holds significant potential for contextual phenotyping and improved decision-making in resource-limited settings.
• This approach can enhance personalized care strategies by identifying nuanced patient subgroups.