Tracing the Heart: An Evidence-Linked Pipeline for Heart-Failure Feature Engineering

Analysis

Why This Paper Matters

Feature engineering in electronic health records (EHR) is a notorious bottleneck, consuming 39-45% of data scientists' time. In heart failure—a condition affecting 6.7 million U.S. adults—the challenge is compounded by fragmented data and the need for guideline-based clinical reasoning. Existing rule-based and LLM-based approaches offer only partial automation, often lacking maintainability and evidence traceability. This paper introduces the Nimblemind Multi-Agent System (nMAS), a pipeline that not only automates feature engineering but also embeds evidence linkage and rubric-based auditing, directly addressing the trust and transparency requirements of clinical AI.

The significance extends beyond heart failure. The methodology—combining multi-agent orchestration with rubric-grounded generation and LLM-based audit—provides a template for other complex, high-stakes domains where feature engineering must be both scalable and defensible. By demonstrating that automated features can improve phenotyping performance, the paper makes a strong case for moving away from manual, error-prone feature construction.

Technical Contributions

• Multi-Agent Architecture (nMAS): A pipeline that decomposes feature engineering into specialized agents, each handling generation, verification, and auditing, ensuring modularity and maintainability.
• Evidence-Linked Generation: Features are explicitly linked to clinical guidelines and source EHR tables, providing provenance and supporting clinical reasoning.
• Rubric-Grounded Scoring: Aggregated features are scored against a rubric, enabling objective quality assessment and consistency.
• LLM-Based Audit: A restricted LLM independently audits features for evidence support and methodological soundness, adding an extra layer of verification.
• Comprehensive Evaluation: The pipeline was tested on 500 dummy patient records from nine EHR source tables, generating 132 structured and 70 rubric-scored features, with structural integrity and provenance checks.

Results

The paper reports concrete performance gains. Adding nMAS-generated features improved held-out AUROC from 0.895 to 0.963 for HFrEF phenotyping and from 0.870 to 0.910 for HFpEF phenotyping. These are substantial improvements, particularly for HFrEF, suggesting that the automated features capture clinically relevant patterns that standard features miss. Additionally, an independent LLM-based rubric assessment scored the features at 81.5% of maximum points, indicating high evidence support and methodological soundness. The features also passed structural integrity and provenance checks, confirming their reliability.

Significance

This work has the potential to reshape clinical AI workflows by reducing the manual burden of feature engineering while improving model performance and auditability. The nMAS approach could be adapted to other disease areas, making it a valuable contribution to the broader AI-for-healthcare field. However, the single-institution dummy data limitation underscores the need for external validation on real-world cohorts. If validated, this pipeline could accelerate clinical research and enable more transparent, trustworthy AI models in healthcare.