Traditional machine learning requires centralizing data for model training, but healthcare data is highly sensitive and regulated by laws like GDPR in Europe and HIPAA in the US. Compliance issues often make transferring this data impossible. Federated Learning, enhanced with privacy-preserving and security techniques, ensures that data remains protected while staying at its original point of storage. ML models are sent to each data location where they are trained locally. Only the model updates or parameters are sent back to a central server where they are aggregated.

Healthcare data is dispersed across many institutions, and individual organizations often lack sufficient data to build effective models. In Federated Learning, since the data does not need to be shared, more institutions might be willing to participate. This increases the available data for training models without compromising privacy.

Using data from a single institution can introduce bias due to demographic, specialist, institutional, machine type, or geographic factors. This can result in models that perform well on training data but poorly on underrepresented groups. Federated Learning trains models by using datasets from different sources, which can help reduce the risk of bias. The aggregated model better captures the diversity of the entire population it serves.

In the case of centralized ML, in order to train a model, the compute and storage resources need to be available on the server. In Federated Learning, the data remains where it was generated. The computing is shared among all the participants, therefore reducing the need for a single powerful infrastructure required to handle the training.