Classification of hemoglobin fractions in the liquid state using Raman spectroscopy combined with machine learning

Identifying hemoglobinopathies is important for the clinical management of many diseases. One of the common techniques to screen hemoglobinopathies is through high-performance liquid chromatography separation followed by UV-VIS detection. Although UV-VIS can quantify the hemoglobin fractions, it is unable to identify them. Here, we use Raman microscopy to generate fingerprint spectra of hemoglobin fractions based on which the fractions can be identified. Five different hemoglobin types are investigated in their liquid state: HbA0, HbS, HbF, HbA1c, and HbA2. Machine learning models based on support vector machines and fully-connected neural networks are optimized to classify these fractions achieving 98.2 & PLUSMN; 0.1% and 98.5 & PLUSMN; 0.3% test F1-score, respectively. In addition, the test accuracy of these two models are 98.2 & PLUSMN; 0.1% and 98.5 & PLUSMN; 0.3%, respectively. Our approach demonstrates the potential of Raman spectroscopy as an identification module in combination with high-performance liquid chromatography. Moreover, this detection approach can be easily miniaturized and integrated with microfluidics.