No abstract

Significance: The creation of fundamentally new approaches to storing various biomaterial and estimation parameters, without irreversible loss of any biomaterial, is a pressing challenge in clinical practice. We present a technology for studying samples of diabetic and non-diabetic human blood plasma in the terahertz (THz) frequency range. Aim: The main idea of our study is to propose a method for diagnosis and storing the samples of diabetic and non-diabetic human blood plasma and to study these samples in the THz frequency range. Approach: Venous blood from patients with type 2 diabetes mellitus and conditionally healthy participants was collected. To limit the impact of water in the THz spectra, lyophilization of liquid samples and their pressing into a pellet were performed. These pellets were analyzed using THz time-domain spectroscopy. The differentiation between the THz spectral data was conducted using multivariate statistics to classify non-diabetic and diabetic groups' spectra. Results: We present the density-normalized absorption and refractive index for diabetic and nondiabetic pellets in the range 0.2 to 1.4 THz. Over the entire THz frequency range, the normalized index of refraction of diabetes pellets exceeds this indicator of non-diabetic pellet on average by 9% to 12%. The non-diabetic and diabetic groups of the THz spectra are spatially separated in the principal component space. Conclusion: We illustrate the potential ability in clinical medicine to construct a predictive rule by supervised learning algorithms after collecting enough experimental data.

In this research, an experimental U87 glioblastoma small animal model was studied. The association between glioblastoma stages and the spectral patterns of mouse blood serum measured in the terahertz range was analyzed by terahertz time-domain spectroscopy (THz-TDS) and machine learning. The THz spectra preprocessing included (i) smoothing using the Savitsky–Golay filter, (ii) outlier removing using isolation forest (IF), and (iii) Z-score normalization. The sequential informative feature-selection approach was developed using a combination of principal component analysis (PCA) and a support vector machine (SVM) model. The predictive data model was created using SVM with a linear kernel. This model was tested using k-fold cross-validation. Achieved prediction accuracy, sensitivity, specificity were over 90%. Also, a relation was established between tumor size and the THz spectral profile of blood serum samples. Thereby, the possibility of detecting glioma stages using blood serum spectral patterns in the terahertz range was demonstrated.