CoChat

A generative AI multi-agent framework with integrated XAI governance for cancer diagnostics: from multi-omics interpretation to lifestyle risk stratification

Open access
Barki, Chamseddine, Chouchen, Mariem, Sediri, Afef, Ceylan, Halil İbrahim, Rahmouni, Hanene Boussi, Muntean, Raul Ioan, El-Seedi, Hesham R., Bragazzi, Nicola Luigi, Dergaa, Ismail · 2026 · Frontiers in Systems Biology

The DOI matches its Crossref record; 58 of the 58 references it cites were found in a registry or an index; 52 of the 76 sources its claims cite were checked for support.

Abstract

Background: Cancer diagnostics is being reshaped by rapid advances in artificial intelligence, yet a persistent gap separates computational performance from clinical trust. Systematic reviews confirm that 83% of XAI studies in oncology excluded clinicians from development or evaluation, 87% lacked rigorous assessment of XAI explanations, and no universally accepted quality metrics for XAI outputs currently exist. Concurrently, generative AI (GenAI) is entering oncology at an unprecedented pace, yet it operates largely without formal interpretability governance. Aim: This review aimed to (i) synthesize the documented gaps in XAI applications for cancer diagnostics from peer-reviewed literature, (ii) evaluate the state and limitations of GenAI in oncological settings, and (iii) propose a conceptual framework of three GenAI-powered, XAI-governed agents designed to address these gaps within a systems biology context. Review Findings: A narrative synthesis of peer-reviewed literature published between 2020 and 2026 across PubMed, Scopus, and Web of Science identified four critical, recurrent gaps: systematic exclusion of clinicians from XAI development, the absence of standardized evaluation metrics, incomplete cross-omics explanations, and the near absence of lifestyle-driven XAI models for cancer risk. GenAI is accelerating in oncology but introduces additional safety concerns, including hallucinations, non-deterministic outputs, and the illusion of transparency through chain-of-thought reasoning. Framework Proposal: Three specialized agents are proposed: a Multi-Omics XAI Agent (MO-XAI Agent) for cross-layer biological data interpretation, a Clinician Trust and Communication Agent (CTC Agent) for structured explanation translation and quality scoring, and a Lifestyle-Driven Cancer Risk Stratification Agent (LRS Agent) for modifiable risk factor analysis. Each agent uses a large language model as the computational backbone with SHAP, LIME, or graph-level XAI methods as governance layers. Conclusion: To our knowledge, this framework represents the first conceptual architecture to integrate agentic GenAI with systematic XAI governance for cancer diagnostics, offering a clinically grounded, biologically interpretable, and regulatorily aligned design specification for future implementation research.

Overview

CoChat Summary

TL;DR

This review synthesizes documented gaps in XAI for cancer diagnostics and GenAI limitations in oncology, then proposes a three-agent GenAI/XAI-governed conceptual framework (MO-XAI, CTC, LRS) for multi-omics interpretation, clinician communication, and lifestyle risk stratification. It is a design specification and research agenda, not an empirically validated system.

Key findings

  • The narrative synthesis identified four recurrent gaps: clinician exclusion from XAI development (83% of XAI cancer studies excluded clinicians; 73% excluded them from system design), absent standardized evaluation metrics (87% lacked rigorous assessment of XAI explanations), incomplete cross-omics explanations, and near absence of lifestyle-driven XAI models for cancer risk.
  • GenAI is accelerating in oncology but introduces safety and interpretability concerns, including hallucinations, non-deterministic outputs, and chain-of-thought reasoning as an illusion of transparency; cited LLM benchmarks include GPT-4 matching pathologist-level error detection (89.5% vs. 88.5%), dermatologist skin lesion classification (84.8% vs. 84.6%), and ovarian cancer staging (97% vs. 88% by radiologists).
  • The proposed framework has three agents: MO-XAI for cross-layer multi-omics interpretation using SHAP/GNNExplainer and LLM pathway interpretation; CTC for explanation translation and quality scoring adapted from CLIX-M; and LRS for lifestyle risk stratification using XGBoost, SHAP, and RAG-supported causal annotation.
  • Multi-agent coordination is cited as improving most likely diagnosis accuracy from 19.65% to 34.11% with GPT-4; validation pathways include TCGA across 33 cancer types and UK Biobank with approximately 500,000 participants.

Methods

  • Structured narrative review, not a systematic meta-analysis, searching PubMed, Scopus, IEEE Xplore, and Web of Science primarily for 2020–2026 publications on XAI, cancer diagnostics, interpretability, SHAP, LIME, Grad-CAM, and multi-omics cancer AI.
  • Inclusion required AI applications in cancer diagnosis, prognosis, or treatment decision support with explainability/interpretability; exclusions covered non-oncology, non-AI, or non-explainable predictive models. Findings were synthesized thematically across XAI methods/modalities, clinical application domain, interpretability evaluation strategies, and reported limitations/gaps.
  • The framework proposal describes three modular agents sharing an LLM backbone and XAI governance layers, coordinated by an orchestration layer that aggregates intermediate outputs into a Unified Clinical Report, with regulatory alignment to EU AI Act and FDA guidance.

Limitations

  • Stated: the work is conceptual and exploratory, includes no empirical datasets for statistical validation, and the framework has not been empirically validated; it should be read as a research agenda requiring multi-institutional collaboration, clinical co-design, implementation, and prospective evaluation.
  • Stated: XAI method choices reflect the current evidence base but may change; fairness requires dedicated research; CTC rubric co-design is needed; LRS causal annotations require validation against Mendelian randomization and randomized lifestyle-intervention trials; economic implications have not been formally evaluated.
  • Evident from the text: no standardized dataset currently exists for CTC Agent validation, no prospective clinical deployment or cost-effectiveness analysis is possible, and the paper relies on literature synthesis and proposed architecture rather than primary data.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3389/fsysb.2026.1892418

Cites

58 of 58 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

FIGURE 1 Conceptual map of documented XAI gaps in cancer diagnostics | Evidence-based gap taxonomy across four dimensions.
FIGURE 1 Conceptual map of documented XAI gaps in cancer diagnostics | Evidence-based gap taxonomy across four dimensions.CC BY · PubMed Central
FIGURE 2 Overview of the architecture of the three-agent Agentic XAI Framework with orchestration layer and clinical output pathways.
FIGURE 2 Overview of the architecture of the three-agent Agentic XAI Framework with orchestration layer and clinical output pathways.CC BY · PubMed Central
FIGURE 3 Conceptual overview of the MO-XAI Agent architecture: multi-omics integration, cross-layer SHAP attribution, and LLM-assisted pathway interpretation.
FIGURE 3 Conceptual overview of the MO-XAI Agent architecture: multi-omics integration, cross-layer SHAP attribution, and LLM-assisted pathway interpretation.CC BY · PubMed Central
FIGURE 4 CTC Agent: explanation translation, quality scoring, and context-specific clinical formatting workflow.
FIGURE 4 CTC Agent: explanation translation, quality scoring, and context-specific clinical formatting workflow.CC BY · PubMed Central
FIGURE 5 LRS Agent: lifestyle-driven cancer risk stratification with causal-annotated SHAP explanations via RAG-augmented LLM.
FIGURE 5 LRS Agent: lifestyle-driven cancer risk stratification with causal-annotated SHAP explanations via RAG-augmented LLM.CC BY · PubMed Central

Verification

Claims the paper makes

39 Valid · 6 Not valid · 16 Not checked · 47 Own results and interpretations

48 more claims · 21 with their cited sources checked

Create a free account to see all 108 claims and to verify the 2 references still open.

See all claims — free account
  • Probably not a substantial claim. 33
  • It may state two separate claims as one. 30

Create a free account to see these claims.

Citations

Cites

Metadata

Authors
Barki, C., Chouchen, M., Sediri, A., Ceylan, H. İ., Rahmouni, H. B., Muntean, R. I., El-Seedi, H. R., Bragazzi, N. L., & Dergaa, I.
Affiliations
  • Tunis University (TN)
  • Research Laboratory of Biophysics and Medical Technologies, The Higher Institute of Medical Technologies of Tunis, University of Tunis El Manar
  • Department of Physical Education and Sports Teaching, Faculty of Sports Sciences, Atatürk University
  • 1 Decembrie 1918 University (RO)
  • Islamic University of Madinah (SA)
  • Saarland University (DE)
  • University of Jendouba (TN)
Corresponding author
Muntean, Raul Ioan
Venue
Frontiers in Systems Biology
Published
2026-08-03
Volume
6
Pages
1892418-1892418
ISSN
2674-0702 (ISSN-L)
OpenAlex
W7172285805
Cited by
1
References
49
Citations by year
2026: 1
FWCI
0.96
Type
journal article
Language
en
Open access
Locations
  • Frontiers in Systems Biology (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • PubMed Central (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (acceptedVersion)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Physical Sciences › Computer Science › Artificial Intelligence › AI in cancer detection
Keywords
generative AI, multi-agent systems, cancer diagnosis, lifestyle risk factors, cancer risk stratification, large language models, SHAP, lime, hallucinations, clinician trust

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.