CPAgents: Agentic Composite Phenotype Generation for Cardiac Disease Association

Technical Summary: CPAgents – Agentic Composite Phenotype Generation for Cardiac Disease Association

Problem Statement
Identifying robust associations between cardiac imaging phenotypes and clinical diseases is critical for population-scale cardiovascular research and risk stratification. However, current Phenome-wide Association Studies (PheWAS) face significant limitations. They predominantly rely on pre-defined, single-variable phenotypes or expert-crafted features. This approach restricts the ability to capture clinically meaningful non-linear effects and cross-phenotype interactions. Furthermore, while automated feature construction exists, existing methods often lack explicit confounder control, numerical safeguards, and the interpretability required for large-scale phenotype discovery in heterogeneous cohorts. The challenge lies in navigating the high-dimensional, redundant, and complex dependency structure of imaging-derived traits without incurring excessive multiple-testing burdens or producing unstable, uninterpretable findings.

Methodology: The CPAgents Framework
The authors propose CPAgents, an iterative, agent-based framework designed to automatically construct and validate interpretable composite phenotypes from base cardiac imaging features. The system operates as a closed-loop coordination of three specialized agents:

1. Statistical Profiling (Analyst Agent):
   Before generating new features, the Analyst systematically profiles the current feature set ($F^{(t)}$) to uncover data characteristics. It generates five statistical reports:

   • Univariate shape assessment (using $\Delta$AIC between linear and spline logistic models) to detect non-linearity.
   • Distributional characterization (skewness, kurtosis, outliers).
   • Pairwise interaction screening via mutual information.
   • Group-effect analysis (intraclass correlation and Cramér's V) to assess safe group-level standardizations.
   • Redundancy clustering based on absolute correlation matrices.
     These diagnostics are compressed into structured recommendations for the next phase.

2. Transformation Generation (Proposer Agent):
   Guided by the Analyst's hints, the Proposer generates new composite feature candidates ($f_{cand}$) using a Large Language Model (LLM). The generation is constrained to a predefined operator space $O$ (Abstract Syntax Trees) to ensure validity and safety. Candidates are produced via three distinct LLM calls with a 2:2:1 ratio:

   • Medical Priors (40%): Features derived from the LLM's internal clinical knowledge (e.g., established physiological ratios).
   • Statistical Transformations (40%): Features addressing specific statistical reports (e.g., applying log1p to heavy-tailed distributions).
   • Exploratory Mathematics (20%): Safe, automated mathematical combinations to discover latent patterns.
   • Safety Mechanisms: The system enforces numerical guardrails, such as adding a smoothing term ($\epsilon = 10^{-9}$) to division operators to prevent zero-division. Failed candidates are re-prompted for repair.

3. Multi-Stage Filtering and Validation (Verifier Agent):
   The Verifier evaluates candidates through a rigorous screening process:

   • Marginal Utility Check: Evaluates isolated $\Delta$AUC.
   • Stability Selection: Uses ElasticNet regularization to penalize spurious correlations.
   • Deduplication: Removes redundant features via statistical clustering and AST-based textual matching.
   • Global Importance: Uses SHAP values from a LightGBM model to rank features.
   • Global Acceptance: Top candidates undergo cross-model validation (SVM, LDA, AdaBoost, MLP) on a hold-out set. A feature set update occurs only if the composite metric $S = 0.5 \cdot \text{AUC} + 0.5 \cdot \text{Recall}$ improves on at least two of the four models. The process employs early stopping after three consecutive rejections.

Key Contributions

1. Agentic Iterative Discovery: Formulates phenotype discovery as an iterative analyze-propose-verify loop, enabling the automatic construction of composite phenotypes within a constrained, safe operator set.
2. Statistics-Informed Composition: Integrates statistical structure analysis (distributions, redundancy, correlations) with principled screening and cross-validated metrics to mitigate selection-induced overfitting.
3. Interpretable and Reproducible Outputs: Generates compact, closed-form formulas with transparent construction traces and evidence summaries, facilitating downstream auditing and replication.

Experimental Results
The framework was evaluated on the UK Biobank cardiac imaging cohort (26,893 participants) across nine clinical disease categories (e.g., hypertension, cardiac disease, COPD, diabetes).

• Performance: Across 72 classifier–disease–metric combinations, the CPAgents variants achieved the top rank in 56 cases, compared to 18 for baseline methods (expert-defined features and MESHAgents). The method consistently outperformed baselines across all nine disease categories.
• Ablation Studies: Removing the Verifier agent caused the most severe performance degradation (AUC dropped from 0.686 to 0.590), highlighting its critical role in filtering unstable candidates.
• Interpretability: The discovered composite phenotypes demonstrated a Silhouette score of 0.413, approximately 1.5 times higher than expert-defined features (0.276), indicating significantly improved cluster separability and class distinction.
• Association Patterns: Heatmaps of disease-phenotype associations showed that composite phenotypes produced more concentrated and consistent association patterns compared to single-variable traits.

Significance and Claims
The paper claims that CPAgents enables the scalable discovery of stronger phenotype-disease associations that go beyond expert-driven feature selection. By automatically constructing compact, clinically interpretable formulas, the framework addresses the limitations of linear modeling and manual feature engineering. The authors emphasize that their approach yields "transparent evidence trails," allowing for the discovery of non-linear effects and cross-phenotype interactions that are robust to cohort heterogeneity. The work positions agentic systems as a viable solution for bridging the gap between high-dimensional imaging data and reliable, interpretable clinical risk stratification.