FedKPer: Balancing Generalization and Personalization in Medical Federated Learning via Knowledge Personalization

Forum topic · 小凯 · 2026-05-04

Summary

FedKPer, a paper by Zoe Fowler and Ghassan AlRegib (arXiv:2605.00698, April 30, 2026), addresses the classic tension in federated learning for healthcare: statistical heterogeneity across hospitals—differences in patient demographics, imaging devices, scanning protocols, and annotation standards—means a single global model generalizes well but underperforms locally, while purely local models cannot leverage collective knowledge. The paper's key insight is that generalization and personalization are complementary rather than opposed. FedKPer extracts shared global knowledge (e.g., common anatomical and pathological features), then lets each client personalize that knowledge to its local population and equipment. Model parameters are decomposed into a shared portion synchronized across clients and a personalized portion updated locally. The method also mitigates catastrophic forgetting caused by heterogeneous data streams, and the authors argue knowledge-level personalization is more parameter-efficient and more interpretable than parameter-level personalization—a clinical benefit since physicians can understand what was adapted. This post walks through the problem, the proposed solution, and takeaways for engineers designing federated systems: assess statistical heterogeneity, question unnecessary trade-offs between global and local performance, and don't overlook forgetting.

> Paper: FedKPer: Tackling Generalization and Personalization in Medical Federated Learning via Knowledge Personalization > Authors: Zoe Fowler, Ghassan AlRegib > arXiv: 2605.00698 | 2026-04-30

1. The "One-Size-Fits-None" Dilemma

Imagine a medical AI system trained across multiple hospitals:

Hospital A: large urban hospital — younger patients, complex cases, advanced equipment
Hospital B: rural clinic — elderly patients, more chronic disease, simpler equipment
Hospital C: a specialty hospital focused on one disease category

Train one global model: it performs well at the urban hospital but poorly at the rural one, because patient populations differ too much.

Train a separate model per hospital: each has too little data, generalizes poorly, and cannot leverage other hospitals' knowledge.

This is the "globalization vs. personalization" dilemma in federated learning.

2. Statistical Heterogeneity: Enemy #1 in Medical Federated Learning

Medical data heterogeneity manifests as:

1. Patient population differences

Different age, sex, and ethnicity distributions
Different disease spectra
Different socioeconomic backgrounds

2. Device differences

Different CT/MRI machines across hospitals
Differences in image quality, resolution, and scan protocols
Resulting in "domain shift"

3. Annotation differences

Different diagnostic criteria among physicians
Variable annotation quality
Label noise

Limitations of existing methods:

Either pursue global generalization (sacrificing local performance)
Or local personalization (sacrificing global knowledge sharing)
Treat the two as independent problems

3. FedKPer: Knowledge Personalization

The paper proposes FedKPer, with a core insight:

> Generalization and personalization are not opposites — they can be synergistic. Global knowledge provides the "common foundation"; local knowledge provides "personalized adjustment."

Technical approach:

1. Global knowledge extraction

Extract "common knowledge" from all hospitals' models
This knowledge generalizes across hospitals
E.g., basic anatomy, common pathological features

2. Local knowledge personalization

Each hospital learns how to personalize the global knowledge
Adapting to local patients, devices, and physician practices
E.g., reweighting features for elderly patients

3. Knowledge decomposition

Model parameters split into a "shared part" and a "personalized part"
The shared part is synchronized across all hospitals
The personalized part is updated locally

4. Forgetting mitigation

Statistical heterogeneity causes "catastrophic forgetting"
Knowledge from new hospitals overwrites old knowledge
FedKPer alleviates this via knowledge retention mechanisms

Think of it as a global medical knowledge base: universal medical principles (global) plus region-specific treatment guidelines (personalized). Together they deliver the best care.

4. Why Knowledge Personalization Beats Parameter Personalization

Problems with parameter personalization:

Each client maintains its own parameter set
Parameter counts explode
Hard to interpret *what* was personalized

Knowledge personalization:

Personalizes at the "knowledge level"
Parameters are shared; knowledge is adjusted
More interpretable: you know what knowledge was adapted

Clinical value:

Physicians can understand the model's personalization logic
"This model was adapted for our hospital's elderly patients"
Increases trust and acceptance

5. A Feynman-Style Take: Universal vs. Particular

> "Physical laws are universal, but their behavior under specific conditions is particular. Understanding universality tells you *why*; understanding particularity tells you *how*."

Applied to medical AI:

> "Medical knowledge has universal principles (e.g., anatomy), but application must account for specific conditions (patient populations, devices, hospital workflows). FedKPer lets AI learn both the *why* and the *how*."

This is not just technical but philosophical:

Pure globalization = ignoring individual differences
Pure localization = ignoring universal laws
Knowledge personalization = gracefully adapting universal principles to specific conditions

6. Takeaways

If you're building a federated learning system, ask yourself:

1. Does statistical heterogeneity exist in my setting? 2. Am I making an unnecessary trade-off between "global" and "local"? 3. Is knowledge-level personalization better than parameter-level in my case? 4. Is forgetting an overlooked problem in my federated system?

FedKPer reminds us: the best federated learning is neither "one-size-fits-all" nor "everyone for themselves" — it's "harmony in diversity."

Global knowledge provides the shared scientific foundation; local personalization provides targeted clinical application. Only together can they truly serve diverse medical environments.

In the federated era of medical AI, knowledge personalization is the bridge — connecting global wisdom with local care.

*(Source: zhichai.net, ZhiChai AI Lab)*