Data Silo

A large urban hospital network and a rural clinic both possess chest X-rays with confirmed cases of a rare lung condition, but neither can share images due to HIPAA compliance and the absence of a Data Use Agreement (DUA).

• Each institution trains on only a few hundred examples
• The resulting diagnostic models suffer from high variance and poor generalization
• Federated learning enables collaborative training while images never leave their respective PACS systems

Within a single academic medical center, the radiology department stores annotated CT scans for liver tumor segmentation while the oncology department maintains separate databases with corresponding biopsy-confirmed malignancy grades.

• Radiology has pixel-level annotations but lacks ground-truth pathology confirmation
• Oncology has confirmed diagnoses but no corresponding imaging features
• The institutional enterprise data warehouse often lacks the schema to join these disparate sources

A healthcare system acquires multiple hospitals, each running different Picture Archiving and Communication Systems (PACS) from vendors like GE, Siemens, or Philips. Despite being under one corporate umbrella, the proprietary DICOM metadata schemas and storage APIs create functional silos.

• Querying across PACS instances requires custom integration middleware
• Annotation formats differ between systems, breaking downstream ML pipelines
• FHIR-based normalization layers are often required before federated training can commence

A pharmaceutical company runs multi-site clinical trials across the EU, the US, and China. GDPR, PIPL, and state-level regulations impose strict data residency requirements that prevent raw imaging data from crossing jurisdictional boundaries.

• MRI scans from German sites cannot be exported to US-based analysis servers
• Chinese hospital data must remain within mainland infrastructure
• Cross-silo federated learning with Secure Aggregation allows the global model to learn from all sites without violating any single jurisdiction's sovereignty laws

A long-established cancer registry holds decades of mammography films in physical archives. While newer digital mammograms feed directly into AI training pipelines, the historical films—containing rare, early-stage malignancy patterns—remain trapped in analog silos.

• Digitization projects face funding and prioritization challenges
• The resulting temporal bias means models are trained predominantly on recent, digitally-native cases
• Synthetic data generation techniques can partially compensate, but the ground-truth historical distribution remains inaccessible

Three radiologists across different institutions use different annotation platforms—one uses 3D Slicer, another a proprietary vendor tool, and the third a custom in-house solution. Each exports segmentation masks in incompatible formats (NIfTI, DICOM-SEG, JSON contours).

• The silo is not the images themselves but the label ontologies and coordinate systems
• Harmonizing these annotations into a unified format for centralized training is labor-intensive
• Federated learning with personalized model heads can accommodate heterogeneous label schemas without full standardization

A federated topology specifically designed to bridge data silos between a small number of reliable institutional participants, such as hospitals. Unlike cross-device FL with millions of unreliable edge nodes, cross-silo assumes each participant holds a large, curated dataset and possesses substantial compute resources.

• Participants are known, identifiable, and have stable network connectivity
• Each silo typically trains on its own hardware before sharing updates
• Enables collaborative diagnostic model training across hospital networks without moving protected health information

The primary statistical challenge when bridging data silos in healthcare. Local datasets at different institutions are almost never independently and identically distributed, reflecting the unique patient demographics, equipment manufacturers, and clinical protocols of each hospital.

• A radiology model trained at a trauma center sees different injury distributions than one at a community clinic
• Label distribution skew: one hospital may have 40% positive cases while another has 5%
• Feature distribution skew: different scanner vendors produce varying pixel intensity profiles
• Causes client drift during federated training, requiring specialized optimization techniques like FedProx

The foundational algorithm that transforms isolated data silos into a unified model. Each participating institution trains locally on its private data, then transmits only model weight updates to a central aggregation server. The server computes a weighted average of all contributions to produce an improved global model.

• Raw data never leaves its originating institution
• Communication occurs in discrete communication rounds
• The global model represents consolidated knowledge from all silos without ever accessing their contents
• Weighting can be proportional to each silo's dataset size to prevent smaller sites from dominating

A cryptographic protocol ensuring that even the central server cannot inspect individual contributions when aggregating updates from multiple data silos. Using secure multi-party computation techniques, the server computes the sum of encrypted model updates while each participant's contribution remains masked.

• Prevents the server from reconstructing any single institution's training data from its gradient updates
• Defends against gradient leakage attacks and honest-but-curious servers
• Each silo encrypts its update with pairwise masks that cancel out during aggregation
• Critical for healthcare consortia where institutional competitive concerns exist alongside privacy requirements

The legal and governance mechanism that formally bridges data silos across institutional boundaries. A DUA is a binding contract specifying the permitted uses, security requirements, and liability terms for collaborative model training without data centralization.

• Defines which model architectures and training objectives are permitted
• Establishes audit trail requirements for all data accesses and model updates
• Specifies data retention and deletion policies for intermediate artifacts
• Forms the governance backbone of any cross-institutional federated learning consortium
• Must align with HIPAA, GDPR, and local data residency requirements

The umbrella term for all distributional differences across data silos that make naive federated training unstable. Encompasses variations in feature representations, label relationships, and class balance across participating institutions.

• Quantity skew: hospitals contribute vastly different volumes of training data
• Label distribution skew: disease prevalence varies by geography and specialty
• Concept shift: the same label may represent slightly different clinical presentations at different sites
• Addressed through personalized federated learning and robust aggregation techniques like FedProx