An inflammation-derived and clinical-based model for ischemic stroke recovery

Abstract

Background Neuroinflammatory responses reflecting disease progression are believed to be closely associated with the severity of prognosis in post-stroke. Purpose This study developed a combined predicted model of inflammation-derived biomarkers and clinical-based indicators using machine learning algorithms for differentiation of the functional outcome in patients with subacute ischemic stroke. Methods Clinical blood samples and patient data from individuals with subacute ischemic stroke were collected at admission. Based on activities of daily living assessments followed by a 3-month recovery, patients were categorized into two groups: those with little effective recovery (LE) and those with obvious effective recovery (OE). Serum samples underwent proteomic testing for initial candidates. Subsequently, multidimensional validation of candidates in models of ischemia-reperfusion at protein and mRNA levels was performed. T-test, Receiver Operating Characteristic (ROC), and LASSO analysis in an additional cohort were performed to confirm the clinical variables and candidate biomarkers in the discriminatory sensitivity and specificity between the LE and OE groups. Finally, models were developed based on candidates in the training dataset and predicted stroke recovery outcomes in another new dataset using ten standard two-categorical variable algorithms in machine learning. Results We identified higher tissue inhibitor metalloproteinase-1 (TIMP1) and LGALS3 levels were positively correlated with the severity of prognosis after ischemic stroke rehabilitation. TIMP1 (AUC=0.904, 0.873) and LGALS3 (AUC=0.995, 0.794) were confirmed to address superior sensitivity and specificity in distinguishing ischemic stroke from healthy control and LE group from OE group. The TIMP1 and Lgals3 expression exhibited an evident increase in microglia following ischemia-reperfusion. In addition, inflammation-derived biomarkers (TIMP1, LGALS3) coupled with clinical-based indicators (HGB, LDL-c, UA) were built in a combined model with random forest to differentiate OE from LE in 3-month follow-up with high accuracy (AUC = 0.8). Conclusion Our findings provided evidence supporting the critical prognostic potential and risk prediction of inflammation-derived biomarkers after ischemic stroke rehabilitation in complementary to current clinical-based parameters.