FedSepsis: A Federated Multi-Modal Deep Learning-Based Internet of Medical Things Application for Early Detection of...

Area of Science:

• Computer Science, Artificial Intelligence, Machine Learning
• Medical Informatics, Health Informatics
• Biomedical Engineering, Medical Devices

Background:

• The Internet of Medical Things (IoMT) offers potential for secure clinical decision support systems by integrating diverse electronic health records (EHRs).
• Deploying IoMT systems, especially on low-end devices, presents usability and computational challenges.
• Early detection of sepsis is critical for patient outcomes but often hindered by data accessibility and processing limitations.

Purpose of the Study:

• To introduce FedSepsis, an application for early sepsis detection using EHRs on low-end computational devices.
• To evaluate the efficacy of federated learning techniques for secure, distributed machine learning in an IoMT context.
• To explore the impact of multimodality and advanced deep learning models on prediction accuracy and earliness.

Main Methods:

• Developed FedSepsis, integrating deep learning for prediction and natural-language processing on EHRs.
• Implemented and analyzed two federated learning techniques for secure distributed training.
• Tested the system on Raspberry Pi and Jetson Nano edge devices, reporting system-level performance metrics (CPU, memory, temperature, network traffic).

Main Results:

• FedSepsis demonstrated satisfactory performance, with federated learning achieving results comparable to centralized approaches with a moderate number of devices.
• Multimodal EHR data significantly outperformed single-modality inputs.
• Generative adversarial neural networks excelled in handling sparse EHR data, achieving an Area Under the Precision-Recall Curve of 96.55%, Area Under the ROC Curve of 99.35%, and prediction earliness of 4.56 hours when combined with multimodality.

Conclusions:

• FedSepsis proves the feasibility and benefit of utilizing IoMT concepts with low-end devices for early sepsis detection.
• Federated learning provides a secure and effective mechanism for distributed machine learning in healthcare applications.
• The combination of multimodality and generative adversarial networks offers a powerful approach for leveraging EHR data, improving diagnostic capabilities in resource-constrained settings.