AEGIS — Research Proposal | Agentic Ecosystem for Guided Intelligence in Sarcoma

01 — Project Summary

Transforming Rare Cancer Diagnostics

Sarcomas represent a rare and heterogeneous group of malignancies characterized by complex diagnostic pathways and significant clinical challenges. These tumors encompass more than 100 distinct histological subtypes, each requiring specialized knowledge of subtype-specific morphological features, molecular markers, and therapeutic approaches.

"More than 40% of initial histological diagnoses are modified upon expert review, with major discrepancies related to histological grade (43%), histological type (24%), and combined grade and subtype classifications (29%). Expert review leads to management changes in 14.2% of patients."

AEGIS proposes to address sarcoma diagnostics through an integrated methodology combining large-scale knowledge extraction with clinical application, leveraging large language models (LLMs) and retrieval-augmented reasoning. The system aims to provide a "digital second opinion" backed by comprehensive data and rigorous uncertainty measures— augmenting rather than replacing the pathologist.

~1,000

Expert-curated sarcoma cases at UMG

~21,000

Literature publications to process

>100

Distinct sarcoma subtypes covered

40%+

Diagnostic discordance rate to address

02 — Conceptual Framework

The AEGIS Agentic System

AEGIS integrates three core innovations: a CodeAct-based agentic orchestrator, interpretable machine learning with Artificial Representative Trees, and conformal prediction for certified uncertainty quantification.

AEGIS Conceptual Framework - Agentic AI system for sarcoma diagnostics showing the diagnostic gap, technical approach with CodeAct framework, and solution with confidence-calibrated diagnoses

Figure 1: AEGIS conceptual framework. (1) Diagnostic gap: >100 sarcoma subtypes lead to >40% revised diagnoses after expert referral. (2) Technical approach: CodeAct-based agentic system orchestrating reasoning across curated sarcoma literature, computational pathology tools, external databases, and interpretable statistical models. (3) Solution: Confidence-calibrated diagnoses with traceable evidence and human-interpretable decision trees.

03 — The Consortium

Partners & Leadership

SP1 — Clinical Curation & Digital Pathology

Knowledge Base & Computational Pathology

University Medical Center Göttingen (UMG)
Department of Pathology, Sarcoma Referral Center

PD Dr. med. Hanibal Bohnenberger
Senior Physician, Consortium Leader

SP2 — Agentic AI for Clinical Decision Support

Knowledge Extraction & Agentic Reasoning

Leibniz University Hannover (LUH)
L3S Research Center

Dr. Michelle Tang
Postdoctoral Researcher

SP3 — Interpretable Modeling & Uncertainty

Statistical Methods & Explainable AI

University Medical Center Göttingen (UMG)
Department of Medical Statistics

Jun.-Prof. Dr. Björn-Hergen Laabs
CAIMed Junior Research Group Leader

Consortium Synergies

AEGIS represents a new model for interdisciplinary eHealth research embedded within the CAIMed network: three early-career researchers combining clinical medicine, artificial intelligence, and statistical methodology around a shared challenge that none could address alone.

SP1

Clinical Pathology

Clinical validation & curation

SP2

Computer Science / AI

Knowledge extraction & reasoning

SP3

Statistics / ML

Interpretability & uncertainty

04 — Expected Results

Proposed Outcomes

1

Agentic Clinical Reasoning Framework

A CodeAct-based system demonstrating autonomous multi-step diagnostic reasoning with complete provenance tracking. The agent orchestrates literature retrieval, clinical data integration, image analysis, and interpretable models through unified tool invocation. The sarcoma-specific retrieval infrastructure will be released as the first open-source, domain-specific knowledge engine for rare tumor diagnostics.

2

Interpretable Diagnostic Models with Uncertainty Guarantees

ART-based models providing human-readable decision logic that distill ensemble predictions into transparent decision trees. Conformal prediction delivers calibrated confidence estimates with ≥90% empirical coverage guarantee. The resulting uncertainty atlas maps diagnostic confidence across the feature space of sarcoma classification.

3

Validated Information Extraction Pipeline

A production-grade methodology for transforming unstructured biomedical literature into structured diagnostic knowledge. Validated on ~21,000 sarcoma publications against expert-curated ground truth, achieving precision/recall ≥85% for core diagnostic entities. Outputs aligned with standard terminologies (SNOMED CT, ICD-O-3, NCIt) and MII data models.

4

Few-Shot Computational Pathology Module

Meta-learning approaches enabling sarcoma subtype classification from whole-slide images with limited training examples—directly addressing the data scarcity barrier excluding rare diseases from deep learning advances. Target: AUROC ≥0.85 for primary diagnostic categories with ≤50 training examples per subtype.

5

Benchmarking & Translational Framework

A comprehensive comparative study (Human vs. Agent vs. Explainable Models) to address the capability of agents. Governance framework for GDPR, EU AI Act, and EU MDR. Multi-center deployment pathway via CAIMed with a translational system usability scale.

05 — Methodology

Technical Approach

🧠

Agentic AI Architecture

LLM-based agent system built on the CodeAct framework, capable of autonomous retrieval, extraction, and synthesis of biomedical knowledge with full provenance tracking and multi-step reasoning.

📊

Retrieval-Augmented Generation

Hybrid RAG infrastructure combining dense semantic embeddings with sparse lexical indexing, re-ranking strategies, and rationale-guided approaches for optimal retrieval from ~21,000 sarcoma publications.

🔬

Few-Shot Deep Learning

Meta-learning algorithms (Prototypical Networks, MAML) enabling classification of rare sarcoma subtypes with limited training examples from whole-slide images.

🌳

Artificial Representative Trees

Interpretable surrogate models distilling complex ensemble predictions into human-readable decision trees while maintaining prediction accuracy and fidelity to ensemble boundaries.

📐

Conformal Prediction

Distribution-free uncertainty quantification with finite-sample coverage guarantees, providing calibrated confidence estimates for high-stakes clinical decision support.

🔗

Standards Alignment

All outputs aligned with SNOMED CT, ICD-O-3, NCIt, ICCR reporting elements, and MII data models ensuring FHIR/OMOP interoperability across clinical systems.

06 — Work Plan

Proposed Work Package Structure

WP1

Data Harmonization & Preprocessing

Lead: SP1 (UMG Pathology) Duration: M1–M24

Transform existing clinical documentation into computationally accessible formats. Terminology standardization, quality assessment, diagnostic uncertainty formalization from tumor board records.

WP2

LLM-Based Retrieval System for Knowledge Extraction

Lead: SP2 (L3S) Duration: M1–M30

Establish production-grade RAG infrastructure. PDF parsing, semantic chunking, hybrid retrieval with re-ranking, LLM generator with citations, validation against human-curated datasets.

WP3

Agentic Framework for Clinical Data Integration

Lead: SP2 (L3S) Duration: M1–M36

Develop agentic orchestration with CodeAct for autonomous multi-step reasoning. API layer for tools, multimodal data integration, human-in-the-loop validation, safety guardrails.

WP4

Interpretable Modeling & Uncertainty Quantification

Lead: SP3 (UMG MedStat) Duration: M6–M27

Develop ART generation and conformal prediction implementations. Feature integration, ensemble training, uncertainty atlas with interactive visualization.

WP5

Computational Pathology Module

Lead: SP1 (UMG Pathology) Duration: M1–M30

Foundation model evaluation, weakly supervised MIL, few-shot implementation with Prototypical Networks/MAML, morphometric extraction, multimodal integration with agentic framework.

WP6

Benchmarking Study & Translational Framework

Lead: All Partners Duration: M30–M36

Three-way comparison: Human Experts vs. Explainable Models vs. AI Agents. Usability assessment (SUS), translational blueprint for EU MDR/AI Act compliance.

07 — Milestones

Key Milestones & Success Criteria

M1 — Month 15

Clinical Archive Accessible

Computationally accessible with standardized terminology; validation framework established

M2 — Month 18

RAG Infrastructure Validated

Operational vector store with optimal retrieval framework validated by human experts

M3 — Month 28

Agentic System Operational

Validated agentic system for sarcoma with safety measures

M4 — Month 24

ART Models Operational

Interpretable uncertainty-aware diagnostic models with ART surrogates

M5 — Month 22

Few-Shot Pathology Validated

Computational pathology module operational with demonstrated few-shot capability

M6 — Month 36

Validation Complete

Benchmarking study completed; translational blueprint finalized with regulatory pathway

08 — Research Environment

CAIMed Network & Advisory Board

CAIMed — Center for AI in Medicine

AEGIS is anchored within CAIMed, Lower Saxony's strategic initiative for medical AI research connecting university medical centers, technical universities, and research institutes across the state. CAIMed maintains affiliations with leading German research infrastructures including the Medical Informatics Initiative (MII) and the German Center for Cardiovascular Research (DZHK).

AEGIS Advisory Board

CAIMed Founding Director

Prof. Dr. Wolfgang Nejdl

L3S Research Center, AI methodology guidance

CAIMed Director UMG

Prof. Dr. Niels Grabe

Computational Pathology, digital pathology infrastructure

Clinical Oversight

Prof. Dr. Philipp Ströbel

Director, Institute of Pathology, Sarcoma Referral Center

Statistical Evidence

Prof. Dr. Tim Friede

Director, Institute of Medical Statistics

Project Management

Martin Zielke

CAIMed Junior Research Group Leader for Collaborative AI Development and Validation in Oncology • AEGIS concept origin • CAIMed network coordination