(696a) A Model-Based Algorithm for Subtyping Patients with Clotting Abnormalities Using Thromboelastogram Response

Pressly, M., University of Pittsburgh
Parker, R. S., University of Pittsburgh
Clermont, G., University of Pittsburgh
Neal, M., University of Pittsburgh Medical Center
Objectives: Early diagnosis of clotting abnormalities, or coagulopathies, would allow more rapid, effective and personalized transfusion support for trauma patients. We present a method that leverages thromboelastogram (TEG) data, a point-of-care assessment of clotting functionality in the form of a tracing, to accomplish this goal. While the TEG returns test results within an hour, our method takes the initial minutes of data collection and assigns patients to coagulopathy subgroups based on a database of simulated TEGs and a deterministic model.

Methods: RapidTEGs were collected from the Study of Tranexamic Acid during Air Medical Prehospital transport (STAAMP) and Prehospital Air Medical Plasma (PAMPer) clinical trials. After eliminating incomplete or discontinuous TEGs, 76% of TEGs remained (104 patients, n=104 tracings). For each, an adjudication was made whether there was insufficient coagulation, defined by a maximum amplitude (MA) of the tracing less than 50 mm (n=15), or excessive lysis, defined by a decline in MA of at least 7.5 percent over 30 minutes (n=5). Parameters of a 3-state ordinary differential equation (ODE) model of TEG response based on mass conservation and reaction-engineering principles were fit to individual tracings. Estimated parameters include the effective initial platelet count, and rates of platelet activation, thrombus growth, and lysis (P0, k1, k2, and k3, respectively). Simulations over a mesh comprised of the four parameters, over their clinically relevant ranges, defined a set of 167,400 reference TEG trajectories. For each of the 104 patient TEGs, segments of the initial 5 or 10 minutes of data were used in the analysis. Sum of squared error (SSE) was computed between all simulated trajectories and individual patient trajectory over the specified time, and the 10 nearest simulated trajectories were retained. Therefore, the dataset for analysis includes 1040 parameter values for the top 10 trajectories for each timespan of the 104 patients. Four different logistic regression (LR) models were developed to predict either outcome (insufficient clotting or high lysis) using parameter sets from nearest trajectories of either 5 or 10 minutes of patient data. LR models were developed and tested on 700 and 340 parameter sets, respectively (each with balanced incidence of insufficient coagulation or high lysis). LR performance was described by and AUC of the ROC curve reported on the test sets.

Results: For parameter values identified from 5 minutes of patient data, LR shows that model parameters can be highly predictive of MA and moderately predictive of lysis. Multivariate LR identified low MA (AUC = 0.906). LR less accurately identified excessive lysis with all parameters (AUC = 0.857). For the parameter values identified using 10 minutes of data, LR shows that model parameters are more highly predictive of MA and less predictive of lysis. LR using all parameters identified low MA (AUC = 0.966) with high certainty, while there was no improvement in the ability to identify excessive lysis (AUC = 0.829). The variability in accuracy for predicting lysis may be associated with the lower incidence in the total population (n=5).

Conclusion: This simulated TEG database structure quickly analyzes subgroups of high risk. This method combines the use of deterministic modeling with data analytics to determine subgroup membership. Results show reliable prediction of problems with MA and a potential to predict lysis early-on. Applying this method could allow the point-of-care RapidTEG to provide preliminary readouts (for example, at 5 and 10 minutes) of coagulopathy; and thereby, decrease time to patient-tailored treatment. This is particularly applicable for use with shock and trauma patients, where timing is critical.