Privacy Enhancing and Scalable Federated Learning to Accelerate AI Implementation in Cross-Silo and IoMT Environments.

Federated Learning (FL) is a machine learning technique that enables to collaboratively learn valuable information across devices or sites without moving the data. In FL, the model is trained and shared across decentralized locations where data are privately owned. After local training, model updates are sent back to a central server, thus enabling access to distributed data on a large scale while maintaining privacy, security, and data access rights. Although FL is a well-studied topic, existing frameworks are still at an early stage of development. They encounter challenges with respect to scalability, data security, aggregation methodologies, data provenance, and production readiness. In this paper, we propose a novel FL framework that supports functionalities like scalable processing with respect of data, devices, sites and collaborators, monitoring services, privacy, and support for use cases. Furthermore, we integrate multi party computation (MPC) within the FL setup, preventing reverse engineering attacks. The proposed framework has been evaluated in diverse use cases both in cross-device and cross-silo settings. In the former case, in-device FL is leveraged in the context of an AI-driven internet of medical things (IoMT) environment. We demonstrate the framework suitability for a range of AI techniques while benchmarking with conventional centralized training. Furthermore, we prove the feasibility of developing a user-friendly pipeline that enables an efficient implementation of FL in diverse clinical use cases.

Neural gradient boosting in federated learning for hemodynamic instability prediction: towards a distributed and scalable deep learning-based solution.

Federated learning (FL) is a privacy preserving approach to learning that overcome issues related to data access, privacy, and security, which represent key challenges in the healthcare sector. FL enables hospitals to collaboratively learn a shared prediction model without moving the data outside their secure infrastructure. To do so, after having sent model updates to a central server, an update aggregation is performed, and the model is sent back to the sites for further training. Although widely applied on neural networks, the deployment of FL architectures is lacking scalability and support for machine learning techniques such as decision tree-based models. The latter, when embedded in FL, suffer from costly encryption techniques applied for sharing sensitive information such as the splitting decisions within the trees. In this work, we focus on predicting hemodynamic instability on ICU patients by enabling distributed gradient boosting in FL. We employ a clinical dataset from 25 hospitals generated based on the Philips eICU database and we design a FL pipeline that supports neural-based boosting models as well as conventional neural networks. This enhancement enables decision tree models in FL, which represent the state-of-the-art approach for classification tasks involving tabular clinical data. Comparable performances in terms of accuracy, precision, recall and F1 score have been reached when detecting hemodynamic instability in FL, and in a centralized setup. In summary, we demonstrate the feasibility of a scalable FL for detecting hemodynamic instability in ICU data, which preserves privacy and holds the deployment benefits of a neural-based architecture.