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

Feature engineering consumes one-third to nearly half of data scientists' time in clinical AI projects, representing a significant bottleneck that slows down research and delays potential clinical applications. The nMAS system addresses this by automating the process for heart failure cases, which affects 6.7 million U.S. adults.

Traceability and auditability are critical for responsible clinical AI: Unlike previous approaches, nMAS links features directly to clinical evidence and guidelines, with an independent AI system verifying that features meet quality standards. This "evidence-linked" approach scored 81.5% on independent assessment of methodological soundness, making it easier for clinicians and regulators to understand and trust the AI's reasoning.

Automated feature engineering significantly improved diagnostic accuracy: Adding the system's aggregated features improved predictive performance from 89.5% to 96.3% accuracy for one heart failure type (HFrEF) and from 87% to 91% for another (HFpEF), suggesting that systematic, guideline-based automation can outperform manual approaches.

Current validation is limited and external testing is essential before clinical deployment: The system was evaluated only on 500 dummy patient records from a single institution. The authors acknowledge that real-world validation across different hospitals and patient populations is needed before this tool could be responsibly used in clinical practice.

Multi-agent AI systems with built-in oversight mechanisms offer a path toward more trustworthy clinical AI: By combining automated feature generation with restricted LLM auditing and rubric-based verification, the system demonstrates how multiple AI components can check each other's work—a model for responsible AI development in high-stakes healthcare settings.