Establishing generalizability of wearable-enabled blood volume decompensation status estimation algorithms using...

Area of Science:

• Biomedical Engineering
• Physiological Monitoring
• Machine Learning in Healthcare

Background:

• Accurate assessment of blood volume decompensation status (BVDS) is critical for trauma care, particularly in prehospital settings.
• Previous BVDS models, developed using large animal hypovolemia data, lacked generalizability across different data collections and sensor hardware.
• Evaluating the generalizability of BVDS estimation algorithms across diverse datasets and sensor implementations is crucial for real-world application.

Purpose of the Study:

• To evaluate and optimize the generalizability of a BVDS estimation algorithm using transfer learning.
• To assess the performance of machine learning models for cross-device BVDS estimation using synchronized physiological signals.
• To establish a data-efficient transfer learning framework for wearable BVDS monitoring in scenarios with limited data.

Main Methods:

• Utilized two controlled-hemorrhage datasets with synchronized electrocardiogram, seismocardiogram, and photoplethysmogram signals.
• Collected data using both a research-grade non-ambulatory monitor and an FDA-cleared wearable patch sensor.
• Implemented and compared three transfer learning strategies: weight freezing/fine-tuning, end-to-end fine-tuning, and gradient boosting model augmentation (e.g., XGBoost).

Main Results:

• Models trained solely on wearable data achieved root mean square errors (RMSEs) of 23.01% (MLP) and 20.95% (XGBoost) for BVDS.
• Fine-tuning with non-ambulatory data reduced RMSEs to 22.79% (MLP) and 18.71% (XGBoost).
• Demonstrated the efficacy of knowledge transfer between datasets for improved BVDS estimation accuracy.

Conclusions:

• Transfer learning significantly enhances the generalizability and accuracy of BVDS estimation algorithms.
• The developed framework offers a practical approach for wearable BVDS monitoring in resource-limited and data-scarce environments.
• This research paves the way for more reliable prehospital trauma assessment using wearable sensor technology.