Machine Learning Analysis of MicroRNA Expression Data Reveals Novel Diagnostic Biomarker for Ischemic Stroke

Ischemic stroke (IS) is one of the leading causes of morbidity and mortality worldwide. Circulating microRNAs have a potential as minimally invasive biomarkers for disease prediction, diagnosis, and prognosis. In this study, we sought to use different machine learning algorithms to identify an optimal model of microRNA by integrating the expression data of pre-selected microRNAs for discriminating patients with IS from controls.

The expression level of microRNAs in the peripheral blood of 50 patients with IS and 50 matched controls were assessed through real-time polymerase chain reaction (qRT-PCR). Machine learning algorithms, including artificial neural network, random forest, extreme gradient boosting, and support vector machine (SVM) were employed via R 3.6.3 software to establish diagnostic models for IS.

The IS group had significantly increased expression levels of miR-19a (P < 0.001), miR-148a (P < 0.001), miR-320d (P = 0.003), and miR-342-3p (P < 0.001) compared with the control group. MiR-148a, miR-342-3p, miR-19a, and miR-320d yielded areas under the receiver operating characteristic curve (AUC) of 0.872, 0.844, 0.721, and 0.673, respectively, with 0.740, 0.940, 0.740, and 0.840 sensitivity and 0.920, 0.640, 0.600, and 0.440 specificity, respectively. Model miR-148a + miR-342-3p + miR-19a had the best predictive value when analyzed via SVM algorithm with AUC, sensitivity, and specificity values of 0.958, 0.937, and 0.889, respectively.

The diagnostic value of the combination of miR-148a, miR-342-3p, and miR-19a through SVM algorithm has the potential to serve as a feasible approach to promote the diagnosis of IS.