💻 computer science

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

The paper introduces GRACE, a novel large-language-model framework that iteratively refines hypotheses and evidence to automatically discover the optimal number of clinical subgroups in heterogeneous, multimodal data without requiring prior specification of cluster counts, as validated across Long COVID and Parkinson's disease cohorts.

Original authors: Jing Wang, Zorina Galis, Tong Zhang, Yiming Luo, Amar Sra, Xing Niu, Jie Shen, Qiaomin Xie, Jeremy Weiss

Published 2026-07-14

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Jing Wang, Zorina Galis, Tong Zhang, Yiming Luo, Amar Sra, Xing Niu, Jie Shen, Qiaomin Xie, Jeremy Weiss

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to sort a massive, messy pile of mystery boxes. Inside each box is a person's health data: how they sleep, how they move, their heart rate, and how they feel day-to-day. Your goal is to figure out which boxes belong to the same "team" (a specific type of illness) without anyone telling you how many teams exist or what they are called.

For a long time, scientists had to guess the number of teams before they started sorting. It was like saying, "I think there are exactly three teams," and then forcing the boxes into those three groups, even if the boxes clearly didn't fit. This paper introduces a new detective tool called GRACE (Generate, Retrieve, Align, Converge, Evaluate) that doesn't need to guess the number of teams. Instead, it figures out the number as it goes, discovering the groups naturally.

The Detective's New Tool: The GRACE Cycle

Think of GRACE as a super-smart robot detective that uses a "hypothesis loop." Here is how it works:

The Guess: The robot starts with a simple idea, like "Maybe sick people get better over time."
The Evidence: It reads a batch of patient records (like a stack of diary entries).
The Debate: The robot asks itself: "Does the evidence support my guess? Do these people look alike, or are they totally different?"
The Move: Based on what it reads, the robot makes one of three moves:
- SPLIT: "These two people look too different; they should be in separate teams."
- MERGE: "These two teams are actually the same; let's combine them."
- COMMIT: "Okay, I'm sure. These are the final teams."

The robot keeps doing this loop—reading, debating, and adjusting—until the teams stop changing. It's like a sculptor chipping away at a block of stone until the true shape appears, rather than trying to force the stone into a pre-made mold.

The Big Wins: What GRACE Found

The authors tested this robot detective on three very different types of health data, and it found hidden patterns that old methods missed.

1. The Long COVID Mystery (13,511 People)
In a huge study of people with Long COVID, the robot found three distinct teams without being told there were three:

The Protected: These people had very mild symptoms and stayed active (averaging 7,197 steps a day).
The Responders: These people started sick but got better over time.
The Refractory: These people stayed very sick, with almost no recovery. Their physical activity collapsed to just 4,140 steps a day.

The robot discovered that the difference between these groups wasn't just about how tired they felt; it was about a specific "autonomic" system in the body. The "Refractory" group had a 25-fold higher chance of having autonomic issues (like POTS) compared to the "Protected" group. Interestingly, the robot noticed that their resting heart rates were actually the same across all groups; the real difference was in how much they could move and exercise.

2. The Parkinson's Puzzle (93 Patients)
Next, the robot looked at foot sensors from people with Parkinson's disease walking around. It found two types of walkers:

The Unstable Walkers: They had shaky steps and took longer to turn around.
The Preserved Walkers: They moved more smoothly.

The cool part? The robot never saw the doctors' scores (like how severe the disease was). It figured out the groups just by looking at the foot sensors. When the authors checked later, the "Unstable Walkers" really did have worse scores on standard medical tests, proving the robot was right.

3. The Blood Sugar Test (16 People)
The robot also looked at continuous glucose monitors (devices that track blood sugar). It found two types of days:

Stable Days: Blood sugar stayed in a safe zone.
Wobbly Days: Blood sugar jumped around wildly.

Even though the robot only saw the sugar numbers, the "Wobbly" days matched up with higher lab test results (HbA1c) that the robot had never seen.

What GRACE Does NOT Do (And What It Rules Out)

It's important to know what this tool doesn't do, because the paper is very clear about its limits.

It doesn't work on everything yet. When the robot tried to sort people with depression using only wrist movement data, it found two groups based on how active they were, but it failed to separate the depressed people from the healthy ones. The paper suggests that for some conditions, a single wearable device just isn't enough to see the whole picture.
It doesn't prove cause-and-effect. The paper found that people with higher Body Mass Index (BMI) before getting sick tended to have worse Long COVID. However, the authors explicitly state this is an association, not a proven cause. They used a special math trick to show that about 57.6% of this link happens because higher BMI leads to worse initial sickness, but there is still a direct effect (about 42%) that happens for other reasons. They are careful to say this doesn't mean BMI causes the disease, just that they are linked.
It doesn't need to be a "black box." Unlike some AI that gives an answer without explaining why, GRACE forces the robot to write down its reasoning and check its work against real data.

How Sure Are We?

The authors are very confident in the stability of their findings for the successful cases.

For the Long COVID groups, if they shuffled the data and ran the test 100 times, the groups stayed almost exactly the same every time (a stability score of >0.97).
For the Parkinson's and blood sugar tests, the groups were also very consistent when the data was split up and re-tested.

However, for the depression and heart-rate-only tests, the groups were less stable. The paper suggests that while the robot found some patterns there, they weren't strong enough to be reliable on their own.

The Bottom Line

The GRACE framework is like a new kind of microscope for health data. It doesn't need to know the answer before it starts looking. It reads the data, argues with itself, and slowly reveals the hidden groups of patients. It has successfully found new ways to sort Long COVID and Parkinson's patients, but the authors warn that it's not a magic wand for every disease yet. It works best when the data is rich and the patterns are clear, and it needs more than just a single sensor to solve the trickiest medical mysteries.

Technical Summary: The GRACE Cycle

Problem Statement
Phenotype discovery—the identification of clinically or biologically meaningful subgroups from data—is essential for precision medicine but faces three persistent bottlenecks:

Unknown Cluster Number ( $K$ ): Conventional clustering algorithms (e.g., k-means, Latent Class Trajectory Modelling) require $K$ to be specified a priori. Heuristics like silhouette scores often fail to transfer across heterogeneous modalities, and in complex chronic diseases, the true number of phenotypes is the very variable researchers seek to learn.
Data Heterogeneity: Clinical cohorts often mix tabular demographics, irregular longitudinal scores, continuous wearable streams, and free-text records. Standard pipelines force these diverse modalities into a single distance metric, often discarding the structural information necessary for interpretability.
Scalability to LLM Context Limits: While Large Language Models (LLMs) excel at synthesizing heterogeneous evidence, they are constrained by context windows. Clinical cohorts often exceed $10^4$ to $10^6$ entities, making it impossible for a single LLM call to process an entire dataset. Naive random batching risks mixing distinct phenotypes, defeating the purpose of clustering.

Methodology: The GRACE Cycle
The authors introduce the GRACE Cycle (Generate hypothesis, Retrieve evidence, Align, Converge, Evaluate), a framework where a hypothesis, an LLM, and an evidence base are iteratively refined until they agree. The framework operates in a closed loop:

Hypothesis Seeding: A working hypothesis seeds the loop.
Evidence Retrieval: The LLM reads a batch of entities under the current hypothesis.
Alignment & Iteration: The LLM confirms, rejects, or partially supports the hypothesis, suggesting feature additions or removals. The hypothesis and feature set are updated.
Convergence: The loop repeats until the hypothesis and evidence align across independent resamples.

To address the unknown $K$ , GRACE wraps this cycle with a Graph-of-Thought (GoT) refinement step:

Seed Clustering: Spectral clustering on a small seed count ( $K_0$ ) generates summary cards for each cluster.
LLM Refinement: The LLM reads these summary cards plus a between-cluster similarity matrix.
Decision Moves: The LLM proposes one of three moves: SPLIT (if internal heterogeneity is detected), MERGE (if clusters are redundant), or COMMIT (if the partition is stable).
Termination: The process iterates until the cluster count stabilizes across two independent random subsamples.

Technical Contributions

Prompt Design: A four-component prompt template integrating:
- Task framing (domain-specific role).
- Serialised entity histories (chronological sequences of observations).
- An explicit active hypothesis statement.
- Structured JSON output (labels, rationale, and the SPLIT/MERGE/COMMIT move).
- Instrumentation: Includes fairness pre-processing and pairwise-comparison batching to ensure the LLM compares similar entities rather than labeling them in isolation.
Data-Feeding Strategy: A $k$ -nearest-neighbour (k-NN) similarity graph compresses large cohorts ( $10^4$ – $10^6$ entities) into context-budget-respecting batches. By presenting an entity alongside its $k$ nearest neighbours, the system preserves the signal-to-noise advantage of comparison-based learning at scale.

Key Results
The framework was validated across three heterogeneous phenotyping problems:

Long COVID Subphenotyping (NIH RECOVER Cohort, $n=13,511$ ):
- GRACE discovered three subphenotypes without specifying $K$ : Protected (low burden, high activity), Responder (intermediate burden, recovery over time), and Refractory (persistently high burden, low activity).
- Mechanism: The subtypes are explained by a single autonomic/post-viral-fatigue axis. The Refractory group showed a 25-fold enrichment in dysautonomia compared to the Protected group.
- Wearable Signals: Separation was driven by a collapse in physical activity (steps/day) rather than heart rate metrics.
- Mediation: A mediation analysis revealed that the association between pre-infection BMI and symptom burden is partially mediated by acute severity (57.6%) but retains a significant direct effect.
- Stability: Bootstrap Jaccard stability exceeded 0.97.
Parkinson's Disease Motor Subtyping (PhysioNet gaitpdb, $n=93$ ):
- GRACE discovered two gait subtypes (Unstable-gait and Preserved-gait) without prior $K$ .
- Validation: The subtypes were externally validated against withheld clinical scores (Timed-Up-and-Go, Hoehn–Yahr stage, UPDRS), showing significant separation ( $p=0.002$ for TUG).
- Generality: This demonstrated the framework's ability to transfer across diseases and sensor modalities (foot-sensor gait) without re-specification.
Additional Wearable Cohorts:
- Type-2 Diabetes (CGM): Discovered two glycemic-day subtypes validated against withheld HbA1c levels.
- Depression (Actigraphy) & Cardiac (Holter HRV): GRACE identified physiologically real structures (e.g., autonomic balance axes) even when these structures did not perfectly predict categorical diagnostic labels, delineating the boundaries of wearable-only subtyping.

Significance
The paper claims GRACE is the first framework where an LLM simultaneously serves as the hypothesis-revision agent, the evidence reader, and the soft cluster-assignment oracle. Its primary significance lies in:

Outputting $K$ : It recovers the number of phenotypes as an output via a verifiable graph-cut criterion rather than treating it as an input hyperparameter.
Scalability: It scales to cohorts of $n=13,511$ using a local-LLM compatible architecture, preserving privacy under strict data-use agreements (e.g., RECOVER).
Domain Agnosticism: It demonstrates that LLM-guided iterative reasoning can replace conventional clustering when the number of phenotypes is unknown, across longitudinal, gait, and metabolic data modalities.
Interpretability: The resulting phenotypes are statistically separable, externally validated, and grounded in coherent clinical mechanisms (e.g., the autonomic gradient in Long COVID).