Machine Learning-Based Rapid Multi-Component Quantification in Danshen Injections Using 1h Nmr

Abstract

A rapid method for the quantitative determination of components in Danshen Injection was developed using one-dimensional proton Nuclear Magnetic Resonance (1H NMR). Three modeling algorithms were evaluated: partial least squares regression (PLSR), extreme learning machine (ELM), and Lasso regression. We also explored different approaches for constructing independent variables, including using the full NMR response, principal component extraction, and variable selection through the Grey Wolf Optimizer (GWO). Model performance was assessed using root mean square error (RMSE), correlation coefficient (R), and residual prediction deviation (RPD), with relative deviation (RD) further employed to evaluate prediction accuracy. Rapid quantitative models for 18 compounds in Danshen Injection were successfully developed, with Lasso regression yielding the best results. The correlation coefficient ranged from 0.8305 to 0.9837, and RPD values were above 1.77. Lasso regression combined with GWO variable selection generally produced higher RPD values, offering a more robust model with better generalization ability. This study demonstrates that by leveraging machine learning to handle complex data, reliable models can be developed from 1H NMR spectra without the need for alignment operations, significantly reducing workload and processing time.

Keywords: Machine learning, 1H NMR, Rapid quantitative model, Multiple components, Traditional Chinese Medicine injection, Danshen Injection