Artificial intelligence algorithms to detect cerebral small vessel disease tomography markers and their association with clinical outcomes in patients with acute ischemic stroke undergone thrombectomy

Background: Although mechanical thrombectomy (MT) has expanded its indications over the last years, establishing a robust benefit even for the elderly, some trials have shown neutral effects regarding this subgroup. One possible explanation for these findings is the vulnerability of the patients treated under these trials, which means these populations were frailer and, accordingly, portended a significant burden of cerebral small vessel disease (cSVD), which may work as a surrogate marker for brain frailty. Objective: To study the influence of the cSVD markers on the baseline CT scan in the functional outcome and develop Artificial Intelligence (AI) algorythms to detect them. Methods: A trained vascular neurologist blinded to clinical details evaluated the CT scans of 351 patients. We focused on markers of cSVD (leukoaraiosis, old lacunes, atrophy), and calculated the Brain Frailty Score (BFS) and modified Small Vessel Disease score (mSVD) (Figures 1 and 2). Then we carried out a multivariable back-stepwise logistic regression and random forest to identify the variables related to disability at 90 days, and other model to analyze the association with Hemorrhagic Transformation. We segmented the images with 3DSlicer and develop an U-Net model to detect leukoaraosis. Results: The group with disability at 90 days was older, had higher NIHSS at admission, more prevalence of diabetes mellitus (DM), higher levels of systolic blood pressure and glucose levels at admission, poorer collaterals, lower rate of MT, and higher rates of severe cSVD markers at baseline CT scan. Figure 3 shows the variables associated with poor outcomes in a back-stepwise logistic regression. The model performance was excellent: AUC=0.962 (Figure 3), with no signs of heteroskedasticity (Figure 4). In other models with HT and Symptomatic HT as dependent variables, cSVD markers also were associated with these outcomes. The random forest model determined the variables importance in these models (Figure5). The U-net model had great performance to detect severe leukoariosis (accuracy=100%, macro-precision=100%, AUC of ROC curve=1.00) Conclusion: Markers of cSVD on baseline CT scans are related to poorer outcomes, such as disability, mortality and symptomatic Hemorrhagic Transformation in patients with LVO and might be one of the main causes of the lack of benefit of MT in elderly patients, highlighting its importance even in acute LVO context. We also developed a AI software with great performance to detect severe leukoaraiosis.