Skip to content

FabioLiberti/FENITH2

BranchesTags

Repository files navigation

FENITH

FENITH

Federated Learning in Dynamic and Heterogeneous Environments for European Digital Healthcare:
From Theoretical Framework to Implementation in Italian Hospitals

Titolo originale (IT)

Federated Learning in Ambienti Dinamici ed Eterogenei per la Sanità Digitale Europea:
Dal Framework Teorico all'Implementazione negli Ospedali Italiani

PhD Thesis — Fabio Liberti
www.fenith.org

License Submodules Papers Last Commit DOI Citations Link Checker
Stars Forks

Table of Contents

Overview

This repository serves as the central hub for the PhD research project on Federated Learning (FL) applied to the European Health Data Space (EHDS). It aggregates all research artifacts — experimental platforms, domain-specific applications, governance frameworks, foundational studies, published papers, and thesis documents — into a single, structured meta-repository.

Research Questions

Question Cluster
RQ1 Technological Framework — How to design Federated Learning architectures capable of handling the high statistical, infrastructural, and participation heterogeneity of hospital nodes, while ensuring scalability, efficiency, and reliability of distributed training? Platforms
RQ2 Multidimensional Governance — Which governance models can balance technological innovation, regulatory compliance (GDPR/EHDS), economic sustainability, and socio-ethical principles, enabling federated trust networks in healthcare? Governance
RQ3 Practical Adoption — What are the main barriers and enabling conditions for the effective adoption of Federated Learning in Italian hospitals, in technological, organizational, and cultural terms? Applications
RQ4 EHDS Interoperability — How can Federated Learning be integrated with international healthcare standards (HL7/FHIR, OMOP/OHDSI) to ensure full compatibility with the European Health Data Space? Platforms · Governance
Research Questions (IT)
Domanda Cluster
RQ1 Framework Tecnologico — Come progettare architetture di Federated Learning in grado di gestire l'elevata eterogeneità statistica, infrastrutturale e di partecipazione dei nodi ospedalieri, assicurando scalabilità, efficienza e affidabilità del training distribuito? Platforms
RQ2 Governance Multidimensionale — Quali modelli di governance possono bilanciare innovazione tecnologica, conformità normativa (GDPR/EHDS), sostenibilità economica e principi etico-sociali, abilitando reti federate di fiducia nel contesto sanitario? Governance
RQ3 Adozione Pratica — Quali sono le principali barriere e condizioni abilitanti per l'adozione effettiva del Federated Learning negli ospedali italiani, in termini tecnologici, organizzativi e culturali? Applications
RQ4 Interoperabilità EHDS — In che modo il Federated Learning può essere integrato con gli standard sanitari internazionali (HL7/FHIR, OMOP/OHDSI) per garantire la piena compatibilità con l'EHDS? Platforms · Governance

See the full architecture diagram for a visual overview.

FENITH Diagram

Fig. 1 — FENITH Research Architecture Diagram

Hierarchical map of the entire FENITH meta-repository, showing the flow from Research Components & Inputs (left) to Research Outputs (right). The diagram is organized into five color-coded layers:

  • Foundations (grey) — DHFLPL2 (seminal paper, MDPI 2024), Heterogeneous FL educational collection, HFEDL_Project, and the private thesis repository (tesi_fl_2025). These feed into the experimental platforms.
  • Experimental Platforms (blue) — Three FL research platforms: flopbg (TensorFlow/React, Bambino Gesù collaboration), BLEKFL2 (PyTorch/Flower, Blekinge collaboration), and FL-EHDS-FLICS2026 (Python, EHDS compliance framework). Each platform connects to specific conference publications on the right.
  • Domain-Specific Applications (orange) — Five applied repositories: Questionnaire_FL (hospital adoption survey), CIDE (OMOP/FHIR business models), CIDE2 (XAI deepfake detection), FedHR5.0 (Industry 5.0 HR), and CRISTAIN2025 (law enforcement FA-FedAvg). Each is mapped to its corresponding conference paper.
  • Governance & Ethics (green) — Three governance-focused repositories: AI-DIGOSA (ethics/regulation tensions), icsis2026 (health policy infrastructure), and ICID2026 (EHDS reference architecture with IDSA/Data Mesh).
  • Research Outputs (right column) — All 10 publications (P-M through IFKAD), thesis documents, survey materials, and the final PhD Thesis & Dissertation Framework (EHDS).

Arrows trace the lineage from foundational work through platforms and applications to published outputs.

FENITH Infographic

Fig. 2 — FENITH Project Infographic

Visual summary of the FENITH meta-repository, designed as a single-page overview with four main sections:

  • Central Hub (top center) — Shows FENITH as the aggregation point linking all research artifacts: platforms, applications, governance frameworks, foundational studies, papers, and thesis documents. Includes the repository URL and website link.
  • Four Clusters (middle row) — Each cluster is presented as a card with its repositories:
    • Platforms (teal): flopbg, BLEKFL2, FL-EHDS-FLICS2026 with their tech stacks and key metrics.
    • Applications (orange): Questionnaire_FL, CIDE, CIDE2, FedHR5.0, CRISTAIN2025 with domain labels.
    • Governance (blue): AI-DIGOSA, icsis2026, ICID2026 with focus areas.
    • Foundations (grey): DHFLPL2, DHFLPL, Heterogeneous_FL, plus private repos.
  • Repository Structure (top right) — Tree view of the directory layout (platforms/, applications/, governance/, foundations/, docs/).
  • Publications (bottom left) — Complete table of all 10 papers with codes, titles/conferences, years, and status. Includes DOI link for the foundational P-M paper.
  • Getting Started (bottom right) — Clone instructions with --recurse-submodules and the private repo access note.
  • Footer — Author, affiliation, license (CC BY-NC-SA 4.0).

At a Glance
FL Algorithms 38 unique across 3 platforms (FedAvg, FedProx, SCAFFOLD, FedNova, MOON, ...)
Datasets 35 total — 5 benchmark, 6 clinical imaging, 19 healthcare (+ 5 shared)
Experiments 6,000+ (FL-EHDS-FLICS2026 alone)
Publications 10 papers — 1 published (21 citations), 3 presented, 2 accepted, 4 submitted
Conferences MDPI, ISM, CRISTAIN, CIDE, ITAIS, FLICS, ICSIS, ICID, IFKAD
Submodules 16 repositories (14 public + 2 private)
Frameworks TensorFlow, PyTorch, Flower, Flask, React

Timeline

gantt
    title FENITH Research Timeline
    dateFormat YYYY-MM
    axisFormat %b %Y

    section Foundations
    Heterogeneous FL study           :done, 2023-10, 2024-06
    P-M MDPI Applied Sciences        :done, pm, 2024-06, 2024-09
    DHFLPL platform (v1)             :done, 2024-03, 2024-09
    DHFLPL2 systematization          :done, 2024-09, 2025-03

    section 2025 Conferences
    P-IT ITAIS 2025                  :done, 2025-03, 2025-06
    P-IS ISM 2025 (Malta)            :done, 2025-06, 2025-11
    P-CR CRISTAIN 2025 (CHItaly)     :done, 2025-06, 2025-10
    P-C1 CIDE 2025 (Ploiesti)       :done, 2025-07, 2025-10
    P-C2 CIDE 2025 (XAI DeepFake)   :active, 2025-07, 2025-10
    Questionnaire FL Italian Hospitals :done, 2024-11, 2025-07

    section Platforms
    flopbg (Bambino Gesù)            :done, 2024-09, 2026-03
    BLEKFL2 (Blekinge)               :done, 2025-07, 2026-03
    FL-EHDS-FLICS2026                :done, 2026-01, 2026-03

    section 2026 Submissions
    FLICS 2026 (Valencia, IEEE)      :active, 2026-01, 2026-06
    ICSIS 2026 (Valencia)            :active, 2026-01, 2026-06
    ICID 2026 (Ulaanbaatar, Springer):active, 2026-01, 2026-05
    IFKAD 2026 (Budapest)            :active, 2026-01, 2026-07

    section Internships
    OPBG — Ospedale Pediatrico Bambino Gesù :crit, done, 2023-10, 2025-03
    BTH — Blekinge Institute of Technology (Sweden) :crit, active, 2025-04, 2026-09

    section Thesis
    FENITH meta-repo                 :active, 2026-03, 2026-09
    PhD Thesis writing               :active, 2026-03, 2026-05
    Thesis delivery                  :milestone, 2026-05-31, 0d
    Thesis defense                   :active, 2026-06, 2026-12
Loading

Datasets

35 datasets across 3 platforms (click to expand)

Benchmark (5) — shared across flopbg & BLEKFL2

Dataset Samples Classes Format
MNIST 70,000 10 28x28 grayscale
Fashion-MNIST 70,000 10 28x28 grayscale
CIFAR-10 60,000 10 32x32 RGB
CIFAR-100 60,000 100 32x32 RGB
SVHN 99,289 10 32x32 RGB

Clinical Imaging (6) — flopbg

Dataset Domain Classes Platform
Brain Tumor MRI Neuro-imaging 4 (glioma, meningioma, pituitary, no tumor) flopbg
ISIC Skin Lesion Dermatology 9 (melanoma, nevus, BCC, ...) flopbg
Chest X-Ray Radiology 2 (normal, pneumonia) flopbg
Diabetic Retinopathy Ophthalmology 5 (no DR — proliferative) flopbg
Skin Cancer Dermatology 2 (benign, malignant) flopbg
Brain Tumor (Alt.) Neuro-imaging 4 flopbg

Healthcare (19) — FL-EHDS-FLICS2026

Primary evaluated (8)

Dataset Domain Samples Compliance
PTB-XL ECG Cardiology 21,799 52 EU sites
Cardiovascular Disease Cardiology 70,000 —
Diabetes 130-US Hospitals Endocrinology/EHR 101,766 ICD-9
Heart Disease UCI Cardiology 920 4 hospitals
Breast Cancer Wisconsin Oncology 569 —
Chest X-Ray Radiology 5,856 DICOM
Brain Tumor MRI Neuro-imaging 7,023 DICOM
Skin Cancer (ISIC) Dermatology 3,297 DICOM

Additionally supported (11)

Dataset Domain Samples Compliance
Stroke Prediction Neurology 5,110 —
CDC Diabetes BRFSS Epidemiology 253,680 —
CKD UCI Nephrology 400 —
Cirrhosis Mayo Hepatology 418 —
Synthea FHIR R4 Synthetic EHR 1,180 FHIR R4
SMART Bulk FHIR EHR Standards 120 FHIR
FHIR R4 Synthetic Synthetic configurable FHIR R4
OMOP-CDM Harmonized Clinical Data Model configurable OMOP v5.4
Diabetic Retinopathy Ophthalmology 35,126 —
Brain Tumor MRI (Alt.) Neuro-imaging 3,264 —
ISIC Skin Lesions Dermatology 2,357 —

FL Algorithms

38 unique algorithms across 3 platforms (click to expand)

Core & Optimization (shared across platforms)

Algorithm Description flopbg BLEKFL2 FL-EHDS
FedAvg Weighted parameter averaging (McMahan et al., 2017) Y Y Y
FedProx Proximal regularization for non-IID data (Li et al., 2020) Y Y Y
SCAFFOLD Control variates for client-drift correction Y — Y
FedNova Normalized averaging for unequal local steps Y — Y
FedDyn Per-client dynamic regularizer Y — Y
FedExP Extrapolation-based dynamic step size (POCS) Y — Y
FedSpeed Proximal term with gradient perturbation correction Y — Y

flopbg Exclusive (5)

Algorithm Description
MOON Model-contrastive loss aligning local/global representations
FedDisco KL-divergence weighted client contributions
FedLPA Layer-wise precision-weighted (posterior) aggregation
DeepAFL Frozen feature layers + ridge regression classifier
FedEL Elastic layer selection with configurable budget

BLEKFL2 Exclusive (16)

Algorithm Description
FedAvg-Optimized Enhanced FedAvg with SGD momentum, cosine annealing, AMP
FedLaS Class-weighted loss with knowledge distillation
FedAvgM Server-side momentum variant
q-FFL Fairness-aware client weighting (Li et al., 2020)
NoiseAwareFL Client filtering by data quality score
RobustAggregation Outlier-robust aggregation
FedBN Local batch normalization (Li et al., 2021)
AdaptiveFL Drift detection with adaptive learning rate
ContinualFL Continual learning with task boundaries
EWC Elastic Weight Consolidation (Kirkpatrick et al., 2017)
MetaLearningFL MAML-based inner/outer optimization
DomainGeneralizationFL Domain-invariant feature representations
TransferFL Transfer learning across domains
RegularizationFL SI / MAS / L2 regularization methods
ReplayBasedFL Experience and generative replay
ArchitecturalFL Progressive networks, PackNet, Piggyback

FL-EHDS-FLICS2026 Exclusive (10)

Algorithm Description
FedAdam Server-side Adam momentum
FedYogi Controlled adaptive learning rate
FedAdagrad Server-side gradient accumulation
Per-FedAvg MAML-based meta-learning personalization
Ditto L2-regularized personal model training
FedLC Logit calibration for label skew
FedSAM Sharpness-aware flat minima optimization
FedDecorr Decorrelation against dimensional collapse
FedLESAM Globally-guided sharpness-aware (ICML 2024 Spotlight)
HPFL Shared backbone + personalized classifiers (ICLR 2025)

Byzantine Resilience (FL-EHDS-FLICS2026)

Defense Description
Krum / Multi-Krum Distance-based robust aggregation
Trimmed Mean Trims outlier gradients before averaging
Coordinate-wise Median Per-coordinate median aggregation
Bulyan Krum + trimmed mean combination
FLTrust Server-guided trust scoring

Repository Structure

FENITH/
├── assets/              # Logo, diagrams, infographics
├── platforms/           # Experimental FL platforms (submodules)
├── applications/        # Domain-specific FL applications (submodules)
├── governance/          # Governance, policy, and ethical frameworks (submodules)
├── foundations/         # Foundational research and educational materials (submodules)
├── docs/
│   ├── thesis/          # Thesis documents and outlines
│   ├── papers/          # Published and submitted papers (see INDEX.md)
│   ├── questionnaire/   # Survey research materials
│   ├── architecture.md  # Mermaid architecture diagram
│   └── README.md        # Documentation guide
├── .github/workflows/   # CI: link checker
├── CITATION.cff         # BibTeX citation for the foundational paper
├── CONTRIBUTING.md      # Contribution guidelines
├── CHANGELOG.md         # Version history
├── SECURITY.md          # Security and responsible disclosure policy
└── LICENSE              # CC BY-NC-SA 4.0

Platforms

Repository Description Stack RQ
flopbg FL research platform — Universitas Mercatorum & Ospedale Pediatrico Bambino Gesù. 12 FL algorithms, 11 datasets (5 benchmark + 6 clinical). Python, Flask, TensorFlow, React RQ1, RQ3
BLEKFL2 FL research platform — Blekinge Institute of Technology (Sweden) & Universitas Mercatorum. 17+ FL algorithms, Flower framework. Python, Flask, PyTorch, Flower RQ1, RQ4
FL-EHDS-FLICS2026 FL + EHDS governance framework. 17 FL algorithms, 19 healthcare datasets, differential privacy, secure aggregation. Python RQ1-2-3-4

Applications

Repository Description Domain RQ
Questionnaire_FL Research questionnaire on FL adoption in Italian hospitals. 34 questions, systematic literature review. Healthcare Adoption RQ3
CIDE Federated Learning Health Data Platform - OMOP/FHIR business models for European digital health research networks (EHDS). Digital Health / Business Models RQ2, RQ4
CIDE2 Explainable Federated Learning for secure telemedicine — privacy-preserving deepfake detection in digital health platforms. Digital Health / XAI RQ1, RQ2
FedHR5.0 Privacy-preserving Federated Learning framework for Human Resource Management in Industry 5.0. HR / Industry 5.0 —
CRISTAIN2025 FA-FedAvg — Forensic-Aware Federated Averaging for law enforcement inter-agency collaboration. Criminal Justice —

Governance

Repository Description Focus RQ
AI-DIGOSA Multidimensional framework analyzing tensions between norms, ethics, and innovation in distributed AI healthcare. Ethics & Regulation RQ2, RQ4
icsis2026 FL as policy data infrastructure for territorial healthcare planning in aging European regions (EHDS). Health Policy RQ3, RQ4
ICID2026 Three-layered reference architecture for deploying FL within EHDS — IDSA, Data Mesh, hierarchical FL. Architecture & Standards RQ1, RQ4

Foundations

Repository Description Type RQ
DHFLPL2 Foundational paper — MDPI Applied Sciences 2024 (21 citations). Systematized evolution of the FL research platform. Seminal · P-M RQ1, RQ3
DHFLPL Original repo referenced in the published paper (v1). Legacy · P-M RQ1, RQ3
Heterogeneous_Federated_Learning Curated collection of educational materials on heterogeneous FL (2020+). Educational RQ1
tesi_fl_2025 Thesis core repository. Private All
HFEDL_Project Heterogeneous FL research project. Private RQ1

Publications

Code Title Venue Year Status Repository
P-M Federated Learning in Dynamic and Heterogeneous Environments: Advantages, Performances, and Privacy Problems MDPI Applied Sciences · DOI 2024 Published (21 cit.) DHFLPL2 · (v1)
P-IS FedHR5.0: Privacy-Preserving HR Management in Industry 5.0 ISM 2025 — Univ. of Malta 2025 Accepted FedHR5.0
P-CR FA-FedAvg: Forensic-Aware Federated Averaging for Law Enforcement CRISTAIN 2025 — CHItaly 2025 Accepted CRISTAIN2025
P-C1 Transforming Clinical Silos into Economic Assets: Business Models for European Digital Health Research Networks CIDE 2025 — Ploiesti 2025 Presented CIDE
P-C2 Explainable Federated Learning for Secure Telemedicine: Privacy-Preserving Deepfake Detection CIDE 2025 — Ploiesti 2025 Presented CIDE2
P-IT AI Distribuita e Governance Sanitaria: Analisi Multidimensionale delle Tensioni tra Norme, Etica e Innovazione ITAIS 2025 2025 Presented AI-DIGOSA
FLICS FL + EHDS Governance Framework: Differential Privacy and Secure Aggregation FLICS 2026 — Valencia · IEEE 2026 Submitted FL-EHDS-FLICS2026
ICSIS FL as Policy Data Infrastructure for Territorial Healthcare Planning ICSIS 2026 — Valencia 2026 Submitted icsis2026
ICID Three-Layered Reference Architecture for FL within EHDS ICID 2026 — Ulaanbaatar · Springer 2026 Submitted ICID2026
IFKAD FedHR5.0: Federated Learning for Knowledge Asset Dynamics in Industry 5.0 IFKAD 2026 — Budapest 2026 Submitted Repo TBD

Getting Started

This meta-repository uses Git submodules to link all component repositories.

# Clone with all submodules
git clone --recurse-submodules https://github.com/FabioLiberti/FENITH2.git

# If already cloned, initialize submodules
git submodule update --init --recursive

# Update all submodules to their latest commit
git submodule update --remote --merge

Note: Some submodules (tesi_fl_2025, HFEDL_Project) are private repositories. Access requires appropriate GitHub permissions.

License

This work is licensed under CC BY-NC-SA 4.0.

Individual submodules may have their own licenses — refer to each repository for details.

Author

Fabio Liberti PhD Candidate — Universitas Mercatorum, Rome, Italy

Website ORCID Google Scholar ResearchGate

Institutions & Collaborations

This research has been conducted in collaboration with the following institutions:

Institution Role Period
Universitas Mercatorum — Rome, Italy PhD Program — Home University 2023 – present
Ospedale Pediatrico Bambino Gesù (OPBG) — Rome, Italy Research Internship — Clinical FL platforms and healthcare datasets Oct 2023 – Mar 2025
Blekinge Institute of Technology (BTH) — Karlskrona, Sweden Research Internship — Heterogeneous FL and international collaboration Apr 2025 – Sep 2026

Acknowledgements

This PhD research was carried out under the supervision of:

  • Prof. Barbara Martini — Universitas Mercatorum (thesis supervisor)
  • Prof. Andrea Mazzitelli — Universitas Mercatorum (co-supervisor)

The author gratefully acknowledges the contributions of:

  • Prof. Davide Berardi — Universitas Mercatorum, for the collaboration on the foundational Federated Learning research.
  • Prof. Alberto Eugenio Tozzi — Ospedale Pediatrico Bambino Gesù, Preventive and Predictive Diseases Research Unit, for the mentorship during the institutional research internship.
  • Prof. Sadi Alawadi — Blekinge Institute of Technology (Sweden), for the mentorship during the international research internship.
  • All the Italian hospitals that participated in the Federated Learning adoption questionnaire.

Keywords

federated-learning · european-health-data-space · ehds · privacy-preserving · differential-privacy · secure-aggregation · digital-healthcare-governance · heterogeneous-environments · dynamic-environments · omop-ohdsi · hl7-fhir · data-mesh · industry-5-0

About

Federated Learning in Dynamic and Heterogeneous Environments for European Digital Healthcare: From Theoretical Framework to Implementation in Italian Hospitals

www.fenith.org/

Topics

healthcare phd-thesis dynamic-environments differential-privacy privacy-preserving federated-learning secure-aggregation data-mesh meta-repository hl7-fhir industry-5-0 ehds digital-healthcare-governance heterogeneous-environments omop-ohdsi european-health-data-space

Resources

Readme

License

View license

Contributing

Contributing

Security policy

Security policy
Activity

Stars

1 star

Watchers

0 watching

Forks

0 forks
Report repository

Releases 25

v0.7.2 — GitHub Pages Actions deployment Latest
Mar 19, 2026
+ 24 releases

Packages

 
 
 

Contributors