Federated Timeline Synthesis: Scalable and Private Methodology For Model Training and Deployment

2025-06-29

AI Summary

To address privacy preservation, cross-institutional scalability, and multi-task generalization challenges in training generative foundation models on distributed electronic health records (EHRs), this paper proposes the Federated Timeline Synthesis (FTS) framework. FTS models patient histories as tokenized patient health timelines (PHTs) and enables decentralized training via local autoregressive Transformer learning coordinated with a global generator—requiring only model weight uploads. Innovatively integrating time-series tokenization, federated learning, and Monte Carlo simulation, FTS supports zero-shot inference, counterfactual reasoning, and early clinical warning. Evaluated on MIMIC-IV, models trained on FTS-synthesized data achieve performance comparable to those trained on real data across five clinical prediction tasks. This validates FTS’s strong privacy guarantees (no raw data sharing), effective longitudinal modeling, and robust scalability across heterogeneous healthcare institutions.

Technology Category

• Machine Learning: Distributed Machine Learning & Federated Learning
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language
• Multiagent Systems: Other Foundations of Multi Agent Systems

Application Category

• User Modeling, Personalization and Recommendation: Federated recommendation systems and personalization
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Systems and Infrastructure for Web, Mobile and WoT: Federated Web and WoT systems, including distributed, federated and edge-based data processing

Abstract

We present Federated Timeline Synthesis (FTS), a novel framework for training generative foundation models across distributed timeseries data applied to electronic health records (EHR). At its core, FTS represents patient history as tokenized Patient Health Timelines (PHTs), language-agnostic sequences encoding temporal, categorical, and continuous clinical information. Each institution trains an autoregressive transformer on its local PHTs and transmits only model weights to a central server. The server uses the generators to synthesize a large corpus of trajectories and train a Global Generator (GG), enabling zero-shot inference via Monte Carlo simulation of future PHTs. We evaluate FTS on five clinically meaningful prediction tasks using MIMIC-IV data, showing that models trained on synthetic data generated by GG perform comparably to those trained on real data. FTS offers strong privacy guarantees, scalability across institutions, and extensibility to diverse prediction and simulation tasks especially in healthcare, including counterfactual inference, early warning detection, and synthetic trial design.

Problem

• Training generative models on distributed timeseries EHR data
• Ensuring privacy and scalability in federated learning
• Enabling zero-shot inference for clinical prediction tasks

Innovation

• Tokenized Patient Health Timelines for data encoding
• Federated training with local model weight sharing
• Global Generator for synthetic data and zero-shot inference