Abstract
Background
Hyperfibrinolysis is strongly associated with early post-injury mortality, but the condition is preventable with timely detection and targeted interventions. Unfortunately, current hyperfibrinolysis diagnostic assays such as thromboelastography (TEG) are time-consuming, making them infeasible for use in real-time treatment strategies to guide adjustable, precise, and personalized interventions. Computational models offer a fast alternative to current practice for ascertaining patient hemostatic state from quickly-measurable protein concentrations. However, few models exist to predict clot strength. Our prior biologically-interpretable, phenomenological, dynamical system TEG model does not often satisfactorily capture hyperfibrinolysis because this feature was missing from training data. Here, we present model improvements that are experimentally-driven and theoretically-grounded to better facilitate the replacement of slow patient viscoelastic clotting measurements, with and without hyperfibrinolysis, by rapid and accurate predictions of TEG parameters
in silico
.
Methods
We created a tPATXA dataset to enable model refinements, and we validated TEG parameter predictions using the published Activation of Coagulation and Inflammation in Trauma (ACIT) dataset, which was excluded from model training. Our tPATXA dataset consists of 310 citrated native (CN) TEG assay profiles that were generated from healthy human donor whole blood samples spiked with tissue plasminogen activator (tPA) and tranexamic acid (TXA) at varying concentrations to provide added phenomenological information about fibrinolysis behavior. We then used ACIT clinical data from 93 trauma patients containing 254 citrated kaolin (CK) and 122 citrated kaolin with heparinase (CKH) TEG assays for model validation. We characterized model performance by: the coefficient of determination
R
2
; percent-error predictions of TEG parameters R-time, K-time, alpha angle, maximum amplitude, and time to maximum amplitude; and absolute errors of TEG parameters Ly30 and Ly60.
Results
Our key model update is to make a previously-constant parameter a time-varying function. Our updated model substantially outperforms the prior model in both datasets. On the tPATXA dataset that both models were trained on and then subsequently predicted to verify competency, the model updates improved
R
2
from 0.9726 to 0.9983. On the ACIT validation dataset that was not used for training, the model updates improved
R
2
from 0.9848 to 0.9993, and reduced TEG parameter prediction variance by over 99%. In addition, the updated model had low prediction errors of 1%–13% for R-time, K-time, alpha angle, maximum amplitude, and time to maximum amplitude, and 0.7%–2.2% for Ly30 and Ly60. A mechanistic interpretation of the new parameter is its capture of the formation and breakdown of a fibrin clot mesh over time.
Conclusions
Our refined model offers higher accuracy, consistency, and biological interpretability than previously available. Our updated model’s modular design supports future integration with literature phenomenological models that predict thrombin dynamics, as well as with externally-added controllers for automation. Thus, this work captures broad clinical coagulation insights and also lays the groundwork for real-time, personalized trauma care via control-theoretic tools.