Horizontal Federated Learning

The fundamental prerequisite for HFL is feature alignment across all participating nodes. Every hospital must collect the same clinical variables—such as blood pressure, heart rate, and lab values—using compatible schemas. This is typically enforced through a Federated Common Data Model (CDM) like OMOP.

• Requirement: Identical feature schemas across all clients
• Example: Three hospitals all recording HbA1c, LDL, and BMI for diabetic patients
• Contrast: Vertical FL shares samples but different features; HFL shares features but different samples

In HFL, each client holds a mutually exclusive set of data subjects. Hospital A's patients do not overlap with Hospital B's patients. This is the natural state of medical data—patients typically receive care at a single institution.

• Key Insight: No entity resolution or record linkage is required between sites
• Benefit: Simplifies privacy compliance since patient identities are not matched across institutions
• Scale: Sample sizes can range from thousands (cross-silo) to millions (cross-device)

HFL is the native topology for the Federated Averaging algorithm, the most widely deployed aggregation strategy. Because local models share identical architectures and feature spaces, their weight matrices are directly element-wise averageable.

• Process: Each client trains locally on its own samples, then uploads model weights to the aggregation server
• Formula: Global weights = weighted average of local weights, proportional to local dataset size
• Advantage: Mathematically straightforward convergence guarantees under IID assumptions

Despite aligned features, HFL suffers from statistical heterogeneity when local data distributions diverge. Hospital A may treat predominantly geriatric patients while Hospital B serves a pediatric population, creating label distribution skew and concept drift.

• Label Skew: Disease prevalence differs dramatically across sites
• Feature Skew: Same lab test may have different normal ranges across demographics
• Mitigation: Requires proximal regularization terms like FedProx or personalized FL layers to prevent model divergence

HFL aggregation typically weights each client's contribution by its local dataset cardinality. A hospital with 100,000 records exerts more influence on the global model than one with 5,000 records.

• Rationale: Larger datasets produce lower-variance gradient estimates
• Risk: Can bias the global model toward the data distribution of the largest institution
• Alternative: Equal weighting or inverse-variance weighting can be used to balance influence across heterogeneous sites

The primary clinical motivation for HFL is building models that generalize across diverse patient populations without centralizing protected health information. A model trained on geographically distributed data learns robust features rather than site-specific artifacts.

• Outcome: Reduced overfitting to any single hospital's demographic biases
• Validation: Requires federated model evaluation where each site tests the global model on its local holdout set
• Regulatory Alignment: Satisfies GDPR data minimization and HIPAA minimum necessary standards

Radiology departments across institutions share the same imaging modalities—CT scans, MRI, and X-rays—but serve different patient populations. Horizontal FL allows them to jointly train a tumor detection model without centralizing sensitive scans.

• Each hospital trains locally on its own PACS data using identical feature extraction pipelines
• Only encrypted gradient updates are transmitted to the aggregation server
• The global model learns from diverse patient demographics, improving generalization
• Proven effective for breast cancer screening and lung nodule classification across geographically dispersed sites

Intensive care units collect the same structured vital signs, lab values, and ventilator parameters for different patient cohorts. Horizontal FL enables collaborative training of early warning systems across hospital networks.

• Features include heart rate, blood pressure, SpO2, and serum lactate—standardized across all ICUs
• Models trained on aggregated data from multiple hospitals detect rare deterioration patterns invisible to single-site datasets
• Federated training preserves the privacy of protected health information under HIPAA
• Demonstrated superior AUROC compared to single-hospital baselines in sepsis prediction

Hospitals using the same electronic health record system with standardized ontologies can collaboratively predict 30-day readmission risk without sharing patient-level discharge summaries.

• Features include diagnosis codes, medication lists, and prior admission history—identical schema across sites
• Horizontal FL aggregates learnings from diverse socioeconomic populations
• Each hospital benefits from patterns observed at peer institutions while maintaining data locality
• Enables proactive discharge planning and targeted follow-up interventions

Digital pathology departments capture whole-slide images at standardized resolutions and stain protocols. Horizontal FL allows multiple labs to jointly train cancer grading models on gigapixel images.

• Identical feature extraction pipelines process H&E-stained slides across institutions
• Federated training captures morphological variability across patient populations
• Addresses the rare disease problem—no single hospital has sufficient cases of uncommon malignancies
• Compliant with GDPR and HIPAA requirements for cross-border research collaborations

Cardiology clinics monitor patients using FDA-cleared wearables that record identical ECG lead configurations. Horizontal FL trains arrhythmia classifiers across clinic populations without uploading raw waveforms.

• Features are standardized: RR intervals, QRS duration, and ST-segment measurements
• On-device training preserves continuous monitoring data privacy
• The global model benefits from diverse arrhythmia presentations across age groups and comorbidities
• Enables personalized alert thresholds while maintaining a robust base model

Pharmacovigilance centers monitor the same adverse event reporting forms with standardized MedDRA coding. Horizontal FL enables collaborative drug safety signal detection across national databases.

• Features include drug names, reaction terms, and patient demographics—identical coding systems
• Federated aggregation detects rare adverse events that individual centers lack statistical power to identify
• Maintains regulatory compliance with data residency requirements across jurisdictions
• Accelerates post-market surveillance without requiring centralized data warehouses