Diabetes mellitus poses a significant challenge in clinical settings, necessitating frequent blood glucose measurements for insulin dosage determination. Conventional invasive methods, such as finger pricking, carry risks of infection and skin callusing. Non-invasive monitoring techniques offer a promising alternative for patients with hyperglycemia or hypoglycemia, enabling regular self-monitoring and advancing diabetes research. However, the accuracy and universality of most current non-invasive methods for measuring blood glucose concentration (BGC) remain inadequate. Achieving clinical credibility requires the elimination of individual discrepancies (IDs) in measurements. This study focuses on enhancing monitoring accuracy to a clinically acceptable level by mitigating the effects of IDs. We conducted a detailed analysis of factors influencing Near-Infrared spectroscopy measurements to reduce prediction error. An artificial neural network with a backpropagation algorithm was employed to predict BGC from the acquired spectral data. Experimental results confirm that our proposed BGC prediction model effectively leverages IDs, achieving performance that meets clinical standards.