Federated Learning for Privacy-Preserving Predictive Analytics in Electronic Health Records: A Multi-Institutional Study on Sepsis Onset

Dr. Chen Wei; Prof. Maria Rodriguez; Dr. David Oppenheimer; Dr. Sarah Al-Farsi

Abstract

Background: Predictive analytics using electronic health records (EHRs) holds immense potential for early detection of critical conditions like sepsis. However, data privacy regulations and institutional silos often hinder multi-institutional collaborations essential for building robust, generalizable models. Federated learning (FL) offers a promising solution by enabling collaborative model training without centralizing sensitive patient data. This study aims to evaluate the effectiveness of FL in developing a privacy-preserving predictive model for sepsis onset, comparing its performance against centrally trained models and single-institution models. Methods: A retrospective cohort study was conducted using de-identified EHR data from four distinct hospitals across three continents. Data included vital signs, laboratory results, and demographic information for adult patients admitted to ICUs. Three models were developed: Model A (traditional centralized data training), Model B (federated learning), and Model C (single-site local training). The primary outcome was the Area Under the Receiver Operating Characteristic curve (AUC) for predicting sepsis onset within 6 hours of ICU admission. Results: The federated learning model (Model B) achieved an AUC of 0.88 (95% CI: 0.87-0.90), performing comparably to the centralized model (Model A: AUC 0.89; 95% CI: 0.88-0.90, p=0.12) and significantly outperforming the average performance of single-site local models (Model C: average AUC 0.82; 95% CI: 0.80-0.84, p<0.001). Conclusion: Federated learning enables the development of high-performance predictive models for sepsis onset while preserving data privacy, making it a viable and ethical approach for collaborative AI in healthcare.

Keywords: federated learning, predictive analytics, sepsis, EHR, privacy-preserving AI

The ability to predict the onset of life-threatening conditions like sepsis using widely available electronic health record data is pivotal for improving patient outcomes and resource allocation. However, the sensitive nature of health data poses significant challenges to collaborative research and development of robust predictive models. Federated learning offers a revolutionary approach by allowing multiple institutions to collaboratively train powerful AI models without ever sharing raw patient data, thereby protecting privacy and facilitating broader epidemiological insights. This study directly addresses this critical balance.

The aim of the study.

To assess the feasibility and performance of federated learning in developing a privacy-preserving, multi-institutional predictive model for early sepsis onset from EHR data, and to compare its efficacy against traditional centralized and single-institution modeling approaches.

Material and methods.

This retrospective cohort study included anonymized EHR data from critically ill adult patients admitted to the ICUs of four hospitals: Peking University First Hospital (China), Hospital Clínic de Barcelona (Spain), Mount Sinai Hospital (USA), and King Faisal Specialist Hospital (Saudi Arabia). The dataset comprised vital signs (heart rate, blood pressure, temperature, respiratory rate), laboratory results (leukocyte count, creatinine, lactate), and demographic information. Patients diagnosed with sepsis within the first 48 hours of ICU admission were identified using Sepsis-3 criteria. Three model types were trained: 1) Centralized Model (data pooled physically), 2) Federated Learning Model (horizontal FL with client-server architecture using FedAvg algorithm, 10 communication rounds), and 3) Local Models (each institution trained independently). All models used a Long Short-Term Memory (LSTM) network architecture. The primary outcome was the Area Under the Receiver Operating Characteristic curve (AUC) for predicting sepsis onset within 6 hours. Secondary outcomes included sensitivity, specificity, and F1-score.

The federated learning model (Model B) achieved an impressive AUC of 0.88 (95% CI: 0.87-0.90), demonstrating robust predictive capability for sepsis onset within the specified timeframe. This performance was statistically comparable to the centralized model (Model A), which yielded an AUC of 0.89 (95% CI: 0.88-0.90, p=0.12 for difference). Crucially, both the federated and centralized models significantly outperformed the average AUC of 0.82 (95% CI: 0.80-0.84) achieved by the single-site local models (Model C) (p<0.001). The federated model also achieved a sensitivity of 82.1% and specificity of 81.5%, closely mirroring the centralized model's metrics. Training time for FL was approximately 20% longer than centralized training due to communication overhead but significantly faster than sequential single-site training.

Federated learning represents a powerful and ethically sound paradigm for developing clinical predictive models from distributed, sensitive patient data. This study unequivocally demonstrates that FL can achieve predictive performance on par with centralized data approaches for sepsis onset prediction, while substantially outperforming models trained on isolated single-institution datasets. Its ability to maintain data privacy without sacrificing model generalizability makes FL an indispensable tool for future multi-institutional AI research in healthcare. Further investigations into optimizing communication efficiency and robustness against data heterogeneity are warranted.

1. McMahan HB, et al. Communication-Efficient Learning of Deep Networks from Decentralized Data. arXiv preprint arXiv:1602.05629. 2017.

2. Seymour CW, et al. Assessment of Clinical Criteria for Sepsis: For the Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3). JAMA. 2016;315(8):762-774.

3. Rieke N, et al. The future of digital health with federated learning. NPJ Digit Med. 2020;3:119.

4. Zhang J, et al. Federated Learning for Predicting In-Hospital Mortality of COVID-19: A Multi-institutional Study. J Med Internet Res. 2021;23(7):e28156.

5. Shickel B, et al. Deep Learning for Sepsis Prediction: A Systematic Review. J Biomed Inform. 2019;112:103623.