Large Cancer Assistant (LCA): orchestration framework for oncology

The Large Cancer Assistant (LCA), introduced in an arXiv paper submitted 7 Jul 2026 by Ghassen Marrakchi and Basarab Matei, is a model-agnostic, post-hoc orchestration framework designed for scalable clinical decision support in oncology. The paper formalizes the LCA as a 7-tuple architecture and describes an ingestion-to-output pipeline that intentionally decouples multimodal data routing from downstream AI inference.

What is the LCA and how does it work?

The LCA is an orchestration layer that standardizes multimodal patient data and routes it to black-box AI models while keeping the orchestration logic independent of those models. The authors introduce the Entry Theory, which uses Geometric Deep Learning to map patient data along structural and medical axes, a Cancer Switching Module to direct flows, and a Standardized Intermediate Payload (SIP) that isolates core AI execution from hospital IT infrastructures.

The paper frames the design around “Algorithmic Impermeability,” a principle that enforces a strict separation between orchestration logic and underlying AI models. The 7-tuple formalization defines the system components and their interactions so the routing, data standardization and failure-safety behaviors remain model-agnostic.

How was the LCA validated?

A proof of concept validated the orchestration logic across four technical scenarios and measured routing and safety behaviors. The PoC executed a nominal flow with negligible orchestration overhead, demonstrated invariant routing projection during AI model swaps, and achieved a 100% recall rate in generating targeted Supplementary Data Requests (SDR) under injected data anomalies.

The authors also verified multi-protocol execution capability and strict failure-safety in these scenarios. The paper contains 22 pages with 6 figures, 8 tables and 9 appendices documenting the PoC and the formal definitions that underpin the architecture.

Why does the SIP matter?

The Standardized Intermediate Payload establishes a clean architectural boundary between ingestion/orchestration and feature inference, enabling downstream interoperability with Electronic Medical Records as an independent path forward. By outputting SIP instead of coupling directly to model outputs, the LCA aims to reduce the fragility that comes from tightly integrated, monolithic AI pipelines.

This separation could simplify model replacement and integration work: the PoC evidence shows the routing projection remained invariant when AI models were swapped, which the authors present as empirical proof of the framework’s model-agnostic claim.

What are the core technical claims?

The paper advances three technical constructs: Entry Theory for multimodal standardization using Geometric Deep Learning; a Cancer Switching Module for dynamic orchestration across structural and medical axes; and the SIP that decouples core AI execution from hospital IT. The LCA’s formal 7-tuple architecture and the principle of "Algorithmic Impermeability" are central claims used to argue that orchestration can remain stable even as underlying AI models change.

What to watch

Look for whether the SIP concept is adopted in integrations with real-world Electronic Medical Records and for any follow-up work that moves beyond the PoC to clinical validation. The authors explicitly position SIP as setting the stage for downstream EMR interoperability as an independent future paradigm; the next concrete milestone will be demonstrations of SIP-based EMR connections or deployments in live clinical workflows.

The arXiv submission includes a DOI link (arXiv:2607.06531) and identifies Ghassen Marrakchi and Basarab Matei as authors. The paper is formatted to Elsevier JBI style and lists 14 references supporting its formalism and design choices.