Pasteurization is a crucial processing method in the food industry to ensure the safety of consumables. A major part of contemporary pasteurization processes involves using flash pasteurizer systems, where liquids are pumped through a pipe system to heat them for a predefined time. Accurately monitoring the amount of heat treatment applied to a product is challenging. This monitoring helps ensure that the correct heat impact (expressed in pasteurization units) is applied, which is commonly calculated as a product of time and temperature, taking achievability of the inactivation of the microorganisms into account. The state-of-the-art method involves a calculation of the applied pasteurization units using a one-point temperature measurement and the holding time for this temperature. Concerns about accuracy lead to high safety margins, reducing the quality of the pasteurized product. In this study, the applied pasteurization level was estimated using regression models trained with NIR spectroscopy data collected while pasteurizing fruit juices of different types and brands. Several conventional regression models were trained in combination with different preprocessing methods, including a novel prediction outlier detection method. Generalized juice models trained with the concatenated data of all types of juices demonstrated cross-validated scores of RMSECV ∼2.78 ± 0.09 and r2 0.96 ± 0.01, while separate juice models displayed averaged cross-validated scores of RMSECV ∼1.56 ± 0.04 and r2 0.98 ± 0.01. Thus, the model accuracy ±10–30 % is well within the standard safety margins.