<p>Raman spectroscopy combined with machine learning techniques is a promising approach for quantitative substance analysis. Online Raman spectrometers have intrinsic limits in sampling circumstances, preventing the utilization of surface-enhanced Raman scattering (SERS) approaches and therefore hindering high-precision predictions for low-concentration analytes. This paper introduces an innovative framework that integrates B-spline fitting for feature extraction with a least squares concentration prediction model, which is improved by hyperparameter optimization using a genetic algorithm (GA). The performance of this framework was carefully evaluated against four alternative GA-optimized prediction models: wavelet transform feature extraction with ridge regression, linear regression neural networks, standalone ridge regression, and polynomial fitting using least squares. Experimental validation included Raman spectral datasets obtained from boric acid and nitric acid solutions throughout 11 concentration gradients (0–500 mg/L) that were evenly dispersed within the designated range. A stratified data partitioning approach, which assigned six concentration levels to the test set, while leveraging the remaining five to create three separate training subsets (3, 4, and 5 concentration levels), was employed. A comparative investigation revealed that the B-spline–least-squares model achieved optimal prediction accuracy when it was trained on four concentration levels, resulting in a mean root-mean-square error (RMSE) of 5.83 mg/L for both analytes. The performance hierarchy revealed that the wavelet transform–ridge regression model (5-level training subset, RMSE = 6.02 mg/L) was the second-best method. Linear regression neural networks, ridge regression, and polynomial– least squares models achieved optimal performance with five training concentrations, yielding mean RMSE values of 7.35, 9.17, and 12.21 mg/L, respectively.</p>