Target informed client recruitment for efficient federated learning in healthcare

Modern machine learning and deep learning methods have been widely incorporated in decision making processes in healthcare in the form of decision support mechanisms. In healthcare, data are abundant but typically not centrally available and, therefore, require some form of aggregation to facilitate training procedures. Aggregating sensitive data poses a significant privacy risk, which is why, both in Europe and the United States, legal frameworks regulate the treatment of such data. Whilst these measures protect the individual behind the data, they pose a significant challenge that results in extensive legal administration related to data sharing efforts. Federated learning (FL) offers a way to mitigate these challenges by allowing to learn models in distributed fashion, eliminating the need to aggregate data for the purpose of training. However, FL comes with a new set of challenges related to communication overhead, client selection and efficiency of the FL training procedure, among others.

In this work, we extend on a previously proposed client recruitment approach by incorporating knowledge on the local hardware such that it becomes possible to recruit a subset of clients for the federation based on the construct of client-level representativeness, which is expressed in terms of the local target distribution divergence, sample size, and the underlying hardware.

We show that, for prominent, medical regression and classification tasks, the recruitment approach yields results that are on par, or better, compared to the central and federated approaches. The proposed approach requires a mere fraction of the data for training and reduces the training time by a factor of 3-4. In addition, we show that excluded clients can still significantly benefit from the resulting federated model through local fine-tuning.

By expressing the representativeness of clients in function of the deviation in the local target distribution, the sample size and efficiency of the underlying hardware, we are able to define a recruitment approach that yields a subset of clients for the federation resulting in significantly reduced training time, without harming predictive performance, whilst improving the privacy preserving characteristics compared to the standard FL and central approaches.