A wavelength selection method for discrete wavelength combinations was developed based on equidistant combination-partial least squares (EC-PLS) and applied to a near-infrared (NIR) spectroscopic analysis of hemoglobin (Hb) in human peripheral blood samples. An allowable model set was established through EC-PLS on the basis of the sequence of the predicted error values. Then, the wavelengths that appeared in the allowable models were sorted, combined, and utilized for modeling, and the optimal number of wavelengths in the combinations was determined. The ideal discrete combination models were obtained by traversing the number of allowable models. The obtained optimal EC-PLS and discrete wavelength models contained 71 and 42 wavelengths, respectively. A simple and high-performance discrete model with 35 wavelengths was also established. The validation samples excluded from modeling were used to validate the three models. The root-mean-square errors for the NIR-predicted and clinically measured Hb values were 3.29, 2.86, and 2.90 g·L⁻¹, respectively; the correlation coefficients, relative RMSEP, and ratios of performance to deviation were 0.980, 0.983, and 0.981; 2.7%, 2.3%, and 2.4%; and 4.6, 5.3, and 5.2, respectively. The three models achieved high prediction accuracy. Among them, the optimal discrete combination model performed the best and was the most effective in enhancing prediction performance and removing redundant wavelengths. The proposed optimization method for discrete wavelength combinations is applicable to NIR spectroscopic analyses of complex samples and can improve prediction performance. The proposed wavelength models can be utilized to design dedicated spectrometers for Hb and can provide a valuable reference for non-invasive Hb detection.