ELECTRICAL IMPEDANCE SPECTROSCOPY CAN BE USED TO DIFFERENTIATE CLOT ANALOGS OF DIFFERENT COMPOSITION (ID 724)
Abstract
Group Name
on behalf of Sensome Collaborators: Ronan Balland, Abdul Barakat, Bruno Carreel, Myline Cottance, Elena Gusarova, Julie Lafaurie, Gor Lebedev
Background And Aims
Despite the uncontested efficacy of endovascular thrombectomy (EVT) to treat large vessel occlusion acute ischemic stroke (AIS), recent data suggests that fibrin/platelet-rich thrombi are more resistant to EVT and present worse revascularization outcomes compared to red-blood-cell-rich (RBC-rich) thrombi, which are associated with a reduced number of recanalization maneuvers and better clinical outcomes. As an alternative to the limited capabilities of pre-stroke imaging, we investigated the capacity of electrical impedance spectroscopy (EIS) to determine thrombus composition.
Methods
Using a custom sensor, we acquired a database of electrical impedance spectra of 109 ex-vivo generated swine blood thrombus analogs of variable composition placed in a blood flow loop. Thrombi were categorized as “white” without any RBC, “mix” containing some RBC and “red” containing mainly RBC. This swine thrombus database was used to train a machine-learning model predicting the category of a thrombus.
Results
The global accuracy of the model is 86% ([80%, 90%] = 95% confidence interval) with high sensitivity and specificity for each category. The confusion matrix (Figure 1) reveals that most errors occur in between predicting white and mix categories since mixed thrombi can have a low percentage of RBC. Differentiating any thrombus from flowing blood can be achieved with a high sensitivity of 95% [90%, 98%]. Applied to a database of ex-vivo generated human blood thrombus analogs the global accuracy of the model is 79% [71%, 86%].
Conclusions
These results demonstrate that EIS can be used to determine AIS thrombus composition, and that swine and human thrombus analogs largely share electrical features.
Trial Registration Number
Not applicable
lesiondetectionwithaconvolutionaldeeplearningmodel_poster/thumbnails/0.png)
