The GRACE Cycle: A General Large-Language-Model Framework for Phenotype Discovery with Unknown Cluster Number

Area of Science:

• Computational biology
• Precision medicine
• Artificial intelligence in healthcare

Background:

• Phenotype discovery is crucial for precision medicine but faces challenges with conventional clustering methods.
• Existing methods require pre-specification of the number of clusters (K) and struggle with heterogeneous, multimodal, or longitudinal data.

Purpose of the Study:

• Introduce the GRACE Cycle, a large-language-model (LLM)-assisted framework for automated phenotype discovery.
• Enable the discovery of biologically meaningful subgroups without a priori specification of the number of clusters (K).

Main Methods:

• The GRACE Cycle framework iteratively refines hypotheses, LLMs, and evidence bases.
• Utilizes Graph-of-Thought (GoT) refinement for discovering K, with LLM proposing SPLIT, MERGE, or COMMIT moves.
• Employs a four-component prompt template and a data-feeding strategy for efficient processing of large cohorts.

Main Results:

• Successfully applied GRACE to Long COVID subphenotyping, identifying three distinct subgroups with high stability.
• Discovered two motor subtypes in Parkinson's disease gait data, validated against clinical measures.
• Demonstrated domain-agnostic applicability across diverse wearable chronic disease cohorts.

Conclusions:

• The GRACE Cycle framework offers a novel, LLM-guided approach to phenotype discovery.
• It overcomes limitations of conventional clustering by discovering K and handling complex data.
• GRACE provides a domain-agnostic alternative for identifying patient subgroups when K is unknown.