Agent-Guided Relational Concept Discovery: Toward Interpretable Surgical Margin Assessment

The study addresses the challenge of surgical margin assessment using Rapid Evaporative Ionization Mass Spectrometry (REIMS) data, which is complicated by the noisy and unlabeled nature of intraoperative data and the black-box nature of deep learning models. The research proposes a novel framework, Agent-Guided Relational Concept Discovery, which aims to improve interpretability and generalization by learning meaningful concepts directly from data without requiring predefined concept labels.

The proposed method incorporates a reasoning agent that refines semantic descriptions of learned concepts and adjusts their weight based on diagnostic relevance. These concepts are grounded using a biochemical knowledge graph to ensure consistency with known metabolic relationships. The framework was tested on Skin and Breast Cancer datasets, showing improved balanced accuracy and sensitivity over baseline models, and demonstrated fewer false positives in a representative intraoperative case.

The study highlights the limitations of current supervised concept-based approaches, which rely on difficult-to-obtain concept annotations in complex workflows. By integrating a reasoning agent and a knowledge graph, the proposed framework addresses these limitations, offering a more interpretable and robust model for surgical margin assessment.

Future work suggested by the authors includes exploring more targeted querying strategies and integrating gene-level information to enhance biological specificity and interpretability. The study acknowledges the reliance on general metabolic database queries as a current limitation and aims to refine this aspect in subsequent research.