Two-dimensional low-field nuclear magnetic resonance approach for the detection of metabolic syndrome in human serum

Metabolic syndrome (MetS) has become a major public health challenge in recent years. Nuclear magnetic resonance, as a nondestructive measurement, has been widely used in medical diagnosis and screening. Compared to conventional magnetic resonance imaging and nuclear magnetic spectroscopy, low-field nuclear magnetic resonance is superior in portability, simplicity, and low-cost. In this study, a novel approach based on low-field magnetic resonance analysis is proposed to detect metabolic syndrome. T-1-T-2 correlation relaxometry is employed to measure human serum. Combining with a classification model based on partial least squares discriminant analysis (PLS-DA), the detection of metabolic syndrome is realized. The classification models based on both T-1-T-2 relaxation correlation data and spectrum are compared. Additionally, classification model based on convolutional neural networks (CNNs) is also evaluated for comparison. The experimental results indicate that the PLS-DA model based on the T-1-T-2 spectrum achieves a better performance. Specifically, the area under curve, accuracy, F1 score, recall and precision of the classification model evaluated on testing group are 0.784, 70.91%, 75.00%, 77.42%, and 72.73%, respectively. On the other hand, the testing results of CNN model trained using T-1-T-2 spectrum are 0.774, 65.45%, 61.22%, 48.39%, and 83.33% for AUC, accuracy, F1 score, recall, and precision. The experimental results reveal the limitation of CNN models on small datasets. In this study, T-1-T-2 correlation relaxometry is employed for the analysis of MetS for the first time, and the feasibility of proposed method has been demonstrated.