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

Timely assessment of blood volume decompensation status (BVDS) is essential for effective trauma intervention, especially in prehospital and resource-limited environments. Prior work has shown that BVDS models can be developed for hypovolemia datasets in large animals (pigs), but the ability of these models to generalize across data collections, sites, and specific implementations of the sensor hardware has not yet been evaluated. In this work, we leveraged a recently collected dataset where a new self-contained wearable device was used for data collection, together with transfer learning, to evaluate and optimize generalizability of the BVDS estimation algorithm. We analyze two controlled-hemorrhage datasets that include synchronized electrocardiogram, seismocardiogram, and photoplethysmogram signals collected using (i) a research-grade, non-ambulatory monitor and (ii) a US Food and Drug Administration (FDA)-cleared wearable patch sensor. We evaluate several machine learning approaches for cross-device BVDS estimation, including multilayer perceptrons (MLPs) and tree-based regressors. To enable knowledge transfer, we implement and compare three transfer learning strategies for the datasets: freezing the network weights and fine-tuning only the decision layer, fine-tuning the network end-to-end without freezing, and augmenting a pretrained gradient boosting tree model (e.g., XGBoost) by adding additional trees. Models trained solely on wearable data achieved root mean square errors (RMSEs) for BVDS, as compared to reference standard values derived from invasive catheter-based pressures, of 23.01% and 20.95%, while fine-tuning with non-ambulatory data reduced errors to 22.79% and 18.71%, respectively for MLP and XGBoost. By establishing a data-efficient transfer learning framework, this work offers a practical pathway toward wearable BVDS monitoring in the field with data and label scarcity by demonstrating the efficacy of knowledge transfer between different datasets collected under controlled hemorrhage conditions.