FedKPer: Balancing Generalization and Personalization in Medical Federated Learning

Forum topic · 小凯 · 2026-05-04

Summary

FedKPer is a federated learning framework proposed by Zoe Fowler and Ghassan AlRegib (arXiv: 2605.00698, 2026-04-30) that addresses the tension between global generalization and local personalization in medical AI. Medical federated learning suffers from statistical heterogeneity: differences in patient populations across hospitals, device and imaging protocol variation causing domain shift, and inconsistent annotation standards. Existing methods either optimize a single global model that underperforms at specific sites, or train purely local models that cannot leverage cross-institution knowledge. FedKPer treats generalization and personalization as complementary rather than opposing goals. Its key idea is knowledge personalization: global knowledge (e.g., anatomy, common pathology) is extracted and shared across clients, while each hospital learns personalized adaptations to local patients, devices, and workflows. Model parameters are decomposed into a synchronized shared part and a locally updated personalized part, with mechanisms to mitigate catastrophic forgetting caused by heterogeneity. The authors argue knowledge-level personalization is more interpretable than parameter-level approaches, letting clinicians understand what was adapted. The forum post frames this as harmonizing universal principles with local conditions, offering guidance for practitioners designing federated systems facing statistical heterogeneity.

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

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: Specialty hospital focused on one disease category

A single global model may perform well at the urban hospital but poorly at the rural clinic, because patient populations differ too much. Fully local models suffer from small datasets and poor generalization, unable to leverage other hospitals' knowledge. This is the classic *globalization vs. personalization* dilemma in federated learning.

2. Statistical Heterogeneity: The Top Enemy of Medical Federated Learning

Medical data heterogeneity manifests in three main ways:

1. Patient population differences — age, sex, ethnicity distribution, disease spectrum, socioeconomic background 2. Device differences — different CT/MRI scanners, image quality, resolution, scanning protocols, causing *domain shift* 3. Annotation differences — varying diagnostic standards across physicians, inconsistent label quality, label noise

Limitations of existing methods:

They either pursue global generalization (sacrificing local performance)
Or pursue local personalization (sacrificing global knowledge sharing)
They treat the two as independent problems

3. FedKPer: Knowledge Personalization

Key insight: Generalization and personalization are not opposites — they can be synergistic. Global knowledge provides a *common foundation*; local knowledge provides *personalized adjustment*.

Technical approach:

1. Global knowledge extraction — Extract cross-hospital common knowledge (e.g., basic anatomy, common pathological features) from all clients' models 2. Local knowledge personalization — Each hospital learns to adapt the global knowledge to its local patients, equipment, and physician habits (e.g., reweighting features for elderly patients) 3. Knowledge decomposition — Model parameters are split into a *shared* part (synchronized across hospitals) and a *personalized* part (updated locally) 4. Forgetting mitigation — Statistical heterogeneity causes catastrophic forgetting where new hospitals' knowledge overwrites old knowledge; FedKPer uses knowledge retention mechanisms to mitigate this

Think of it as a global medical knowledge base: universal medical principles (global) plus region-specific clinical guidelines (personalized). Combining both yields the best care.

4. Why Knowledge Personalization Beats Parameter Personalization

Parameter personalization problems:

Each client maintains its own parameter set — parameter count explodes
Hard to interpret *what* was personalized

Knowledge personalization advantages:

Personalization happens at the *knowledge level* while parameters are shared
More interpretable: you know *which knowledge* was adapted

Clinical value: Physicians can understand the model's personalization logic — "this model was adapted for our hospital's elderly patients" — increasing trust and adoption.

5. Universal vs. Particular

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

The same applies to medical AI: medical knowledge has universal principles (e.g., anatomy), but application must account for specific conditions (patient populations, equipment, hospital workflows). FedKPer lets AI learn both the *why* and the *how*.

Pure globalization = ignoring individual differences
Pure localization = ignoring universal regularities
Knowledge personalization = elegantly adapting universal principles to specific conditions

6. Takeaways for Practitioners

If you are building a federated learning system, ask:

1. Does statistical heterogeneity exist in my scenario? 2. Am I making an unnecessary trade-off between globalization and personalization? 3. Is knowledge-level personalization better than parameter-level personalization? 4. Has the forgetting problem been neglected in my federated setup?

**The best federated learning is neither "one-size-fits-all" nor "everyone for themselves" — it is *harmony in diversity*.** Global knowledge provides a shared scientific foundation; local personalization provides targeted clinical application. In the federated era of medical AI, knowledge personalization is the bridge connecting global intelligence and local care.

--- *Translated and edited from a zhichai.net forum post.*