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

Phenotype discovery-the data-driven identification of clinically or biologically meaningful subgroups-is fundamental to precision medicine, but conventional clustering methods require the number of clusters $K$ to be specified a priori and struggle with heterogeneous, multimodal, or longitudinal data. We introduce the GRACE Cycle (Generate hypothesis, Retrieve evidence, Align, Converge, Evaluate), a general large-language-model (LLM)-assisted framework for phenotype discovery in which a hypothesis, an LLM, and an evidence base are iteratively refined until they agree. The framework discovers $K$ as an output through Graph-of-Thought (GoT) refinement, in which an LLM reads per-cluster summary cards plus a between-cluster similarity matrix and proposes one of three moves-SPLIT, MERGE, or COMMIT-over a spectral-clustering seed. Two technical contributions enable scale: (i) a four-component prompt template integrating pairwise comparison, fairness pre-processing, and structured JSON output, and (ii) a data-feeding strategy that compresses cohorts of ${10}^4-{10}^6$ entities into context-budget-respecting batches via $k$ -nearest-neighbour graph sampling. We validate GRACE across three heterogeneous phenotyping problems: (1) longitudinal Long COVID subphenotyping in the NIH RECOVER cohort $(n=\mathrm{13,511})$ , where GRACE recovers three clinically distinct subphenotypes (Protected, Responder, Refractory) with bootstrap Jaccard stability $>0.97$ that are explained by a single autonomic/post-viral-fatigue axis (a 25-fold dysautonomia gradient, dysautonomia adjusted $OR=13.4$ ) and an accompanying collapse of wearable-measured physical activity; (2) motor subphenotyping of Parkinson's disease from foot-sensor gait wearables (PhysioNet gaitpdb, $n=93$ ), where GRACE discovers two gait subtypes without specifying $K$ that are externally validated against withheld Timed-Up-and-Go $(p=0.002)$ , Hoehn-Yahr stage $(p=0.03)$ , and age; and (3) additional open wearable chronic-disease cohorts processed with the identical pipeline. Across domains, GRACE converges without prior knowledge of $K$ , demonstrating that LLM-guided iterative reasoning offers a domain-agnostic alternative to conventional clustering when the number of phenotypes is unknown.