Identification of moyamoya disease based on cerebral oxygen saturation signals using machine learning methods

Moyamoya is a cerebrovascular disease with a high mortality rate. Early detection and mechanistic studies are necessary. Near‐infrared spectroscopy (NIRS) was used to study the signals of the cerebral tissue oxygen saturation index (TOI) and the changes in oxygenated and deoxygenated hemoglobin concentrations (HbO and Hb) in 64 patients with moyamoya disease and 64 healthy volunteers. The wavelet transforms (WT) of TOI, HbO and Hb signals, as well as the wavelet phase coherence (WPCO) of these signals from the left and right frontal lobes of the same subject, were calculated. Features were extracted from the spontaneous oscillations of TOI, HbO and Hb in five physiological activity‐related frequency segments. Machine learning models based on support vector machine (SVM), random forest (RF) and extreme gradient boosting (XGBoost) have been built to classify the two groups. For 20‐min signals, the 10‐fold cross‐validation accuracies of SVM, RF and XGBoost were 87%, 85% and 85%, respectively. For 5‐min signals, the accuracies of the three methods were 88%, 88% and 84%, respectively. The method proposed in this article has potential for detecting and screening moyamoya with high proficiency. Evaluating the cerebral oxygenation with NIRS shows great potential in screening moyamoya diseases. Moyamoya is a cerebrovascular disease with a high mortality rate. In 64 patients with moyamoya disease and 64 controls, we studied the signals of the cerebral tissue oxygen saturation index and the changes in oxygenated and deoxygenated hemoglobin concentrations with the help of near‐infrared spectroscopy. Features were extracted from the wavelet transforms and the wavelet phase coherence of the signals from both frontal lobes of the same subject for machine learning to screen moyamoya diseases. The highest accuracy reaches 88%.