Skip to content

CPAgents: Agentic Composite Phenotype Generation for Cardiac Disease Association

🕒 Published (v1): 2026-06-26 15:20 UTC · Source: Arxiv · Venue: MICCAI 2026 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

CPAgents is an agentic framework that automatically constructs and validates interpretable composite cardiac imaging phenotypes via an iterative Analyst–Proposer–Verifier loop. On the UK Biobank cohort (26,893 subjects, 9 diseases) it outperforms expert-defined and MESHAgent baselines across 56/72 classifier–disease–metric combinations.

Problem

Current phenome-wide association studies (PheWAS) for cardiac imaging rely on pre-defined single-variable phenotypes or expert-crafted features, which limits their ability to capture non-linear effects, cross-phenotype interactions, and group-dependent distributions. Manual feature engineering does not scale to high-dimensional, population-scale imaging data, and naively expanding the feature space inflates the multiple-testing burden and introduces instability.

Method

CPAgents formulates composite phenotype discovery as a closed-loop, three-agent iterative search over a constrained Abstract Syntax Tree (AST) space:

  • Analyst profiles the current feature set via five statistical reports (non-linearity via \(\Delta\)AIC between linear/spline logistic models, distributional shape, pairwise interaction screening via mutual information, group-effect ICC/Cramér's V, redundancy clustering). An LLM compresses these into per-feature structured recommendations.
  • Proposer generates \(C = \min(|F^{(t)}|, C_{\text{max}})\) new AST-based candidates via three parallel LLM calls in a 2:2:1 ratio — medical priors (LLM clinical knowledge), statistical transformations (addressing the Analyst's reports), and exploratory mathematics. All expressions are constrained to a predefined operator set \(O\) (non-linearities, arithmetic interactions, group-aware reductions) with numerical safety guardrails (e.g., \(\epsilon = 10^{-9}\) smoothing on division). Failed ASTs are repaired up to twice.
  • Verifier applies multi-stage filtering: marginal utility (\(\Delta\)AUC), ElasticNet stability selection (50 bootstrap runs, \(\ell_1\)-ratio=0.5, \(C=0.5\)), deduplication (correlation + AST text matching), and SHAP-based global importance ranking.
  • Global acceptance: Surviving candidates are accepted if the composite metric \(S = 0.5\cdot\text{AUC} + 0.5\cdot\text{Recall}\) improves on at least 2 of 4 classifiers (SVM, LDA, AdaBoost, MLP). Early stopping after \(P=3\) consecutive rejections.

The LLM backend is DeepSeek-V3.2 via LangChain. Discovery is restricted to the training set; final evaluation on a held-out test set.

Key Contributions

  • Agentic iterative phenotype discovery: formulate composite phenotype generation as an analyse–propose–verify loop coordinated by specialised LLM agents.
  • Statistics-informed composition and validation: principled screening via cross-validated metrics, stability selection, and ensemble agreement to mitigate selection-induced overfitting.
  • Interpretable and reproducible outputs: each composite is a compact closed-form formula with a clear AST construction trace and cross-validated evidence trail.

Results

  • Ranking over 72 classifier–disease–metric combinations: CPAgents variants (GPT-5.1 and DeepSeek V3.2 backends) occupy the top rank in 56 cases versus 18 for baselines (expert-defined and MESHAgent features). They appear in the top two in 51 settings (vs. 33 and 31 for baselines).
  • Full framework performance: average AUC \(0.686 \pm 0.076\), Recall \(0.643 \pm 0.069\) across all diseases and classifiers.
  • Ablation — Verifier is critical: removing the Verifier drops AUC to \(0.590 \pm 0.006\) and Recall to \(0.544 \pm 0.012\); removing the Analyst, Feedback loop, or SHAP ranking causes smaller degradations.
  • Cluster separability: composite phenotypes achieve a Silhouette score of 0.041 vs. 0.029 for raw features (~41% relative improvement), and 0.041 vs. 0.027 for expert features.
  • Disease–phenotype association patterns: composite phenotypes produce more concentrated, consistent association heatmaps compared to expert-defined features, especially for cardiac disease groups.

Limitations

  • Evaluated only on cardiac imaging (UK Biobank CMR cohort) — single modality, single dataset.
  • Computational cost is bounded by iteration budget \(T\) and early-stopping patience \(P=3\), but per-iteration cost scales with the candidate count \(C\).
  • Quality of generated candidates depends on LLM medical priors — may not generalise to domains with less established clinical knowledge.
  • AST search space is constrained to a predefined operator set \(O\); novel operators require expert extension.
  • Relies on a fixed set of four base classifiers for the ensemble acceptance gate — different classifier panels might yield different discovered phenotypes.

Relevance to Agentic AI / LLM Agents

CPAgents is a direct demonstration of LLM-driven scientific discovery through multi-agent coordination. The Analyst–Proposer–Verifier pattern is a clean instantiation of the "observe–propose–verify" loop that characterises many agentic AI systems for hypothesis generation and data science. For researchers tracking this topic, CPAgents shows how domain-specific agents with constrained tool use (AST operators, statistical tests) can outperform both expert-driven baselines and generic LLM agents (MESHAgent) on structured feature engineering — a task that bridges data science automation and clinical discovery.