National Taiwan University Hospital said in 2026 that it had built a federated-learning platform with partner hospitals to validate medical AI while keeping each institution's data on site. The hospital's clinical AI validation center lists Fu Jen Catholic University Hospital and Min-Sheng General Hospital among its partners and says the platform is designed for different validation needs without moving data out of the hospitals.

The project is one example of a broader approach to training artificial intelligence across organizations that cannot readily pool sensitive records. Federated learning changes where the computation occurs, but it does not by itself resolve privacy attacks, uneven data quality or responsibility for the resulting model.

How federated learning works

Researchers introduced the modern federated-learning framework at the 2017 International Conference on Artificial Intelligence and Statistics. The paper described a shared model trained by aggregating updates computed on distributed devices, leaving the underlying training data where they originated. Its experiments focused on mobile devices rather than hospitals, but the same architecture can connect institutions that hold separate datasets.

A coordinating server sends a copy of the current model to participating sites. Each site trains that copy on its local records and returns an update; the coordinator combines the updates into the next version of the shared model. The US National Institute of Standards and Technology (NIST) says this design allows participants to send model updates instead of data, which remain inside each organization.

A medical worker reviews digital information on a tablet (illustrative image)

Why hospitals are testing it

Hospital records are fragmented across institutions and subject to legal, ethical and operational controls. Centralized training requires the institutions to copy records into a common repository, creating another system that must govern access, security and retention. Federated learning can reduce that movement, although every participating site still needs compatible data definitions, computing capacity and an agreed training protocol.

Taiwan's Personal Data Protection Act supplies the legal context for the National Taiwan University Hospital project. Article 6 of Taiwan's law restricts the collection, processing and use of medical records, healthcare data, genetic data and physical-examination records, while specifying six legal bases under which those activities may occur. Federated learning is therefore a technical control within a wider compliance process, not an exemption from the law.

An engineer reviews an AI training interface and network diagram on computer screens (illustrative image)

A four-hospital study in France

A peer-reviewed study has shown how the approach can work across hospital boundaries. Researchers reported in Nature Medicine in January 2023 that they trained models across four French hospitals to predict how triple-negative breast cancer would respond to preoperative chemotherapy, with patient data remaining behind each hospital's firewall.

The study used pathology images and clinical information from 650 patients. Collaborative training improved performance over local models in the proof-of-concept study, but the authors did not present the system as a deployed clinical service. Access to the underlying patient datasets also remained restricted and required approval from the ethics committee at each center.

A hospital team discusses a multi-institution data and AI project in a meeting room (illustrative image)

Privacy does not end at the hospital firewall

Model updates reflect the records used to produce them. NIST warns that some attacks can recover information about training data from those updates, while other attacks can infer information from the completed model. Defenses include secure aggregation, which limits what the coordinator can see, and differential privacy, which adds controlled noise to reduce the information attributable to individual records.

Those protections bring trade-offs. A 2024 review of privacy preservation in healthcare federated learning found that exchanged information can leak institutional data and surveyed defenses including differential privacy, homomorphic encryption and secure multiparty computation. Hospitals must define the threat model before selecting controls, because no single technique addresses every attack or trust arrangement.

Uneven hospital data can weaken the model

Participating hospitals rarely collect interchangeable data. Patient populations, disease prevalence, scanners, laboratory practices and coding systems may differ, producing what researchers call non-identically distributed data. A global model can perform well on the largest contributors while missing patterns found at a smaller or more specialized hospital.

A 2024 systematic review covering 89 healthcare federated-learning papers identified class imbalance, missing data, distribution shifts and nonstandard variables among the field's recurring problems. The review also found methodological weaknesses across the literature, limiting how confidently results can be compared or generalized.

A screen displays an abstract visualization of encryption and data protection (illustrative image)

Governance remains a separate task

A hospital consortium still has to decide who may join, which records are eligible, how updates are audited and who responds when performance deteriorates. Contracts and technical controls must also cover a participant's withdrawal, security incidents and changes to the model after validation. Keeping raw records local narrows one category of exposure; it does not assign accountability.

The National Taiwan University Hospital page describes a platform for cross-hospital validation and a coordinated ethics-review process. It does not provide a nationwide adoption count, a mandatory rollout schedule or a published evaluation of clinical outcomes. APPI News could not verify how many hospitals outside research and validation projects use federated learning in routine clinical care at the time of writing.

Frequently asked questions

Does federated learning anonymize patient records?
No. It changes the training architecture so that records can remain with the institution that holds them. Hospitals still need access controls, legal authority, security measures and protections against information leaking through updates or the final model.

Is it the same as sharing de-identified records?
No. A centralized project transfers copies of records to a common location after applying its chosen de-identification process. A federated project trains locally and exchanges model updates, though those updates can still carry privacy risk.

Does a federated model work equally well at every hospital?
Not necessarily. Differences in populations, equipment and data collection can produce uneven performance. Each institution needs local validation and ongoing monitoring before relying on a shared model.