Recently, a research team at the Hefei Institutes of Physical Science of the Chinese Academy of Sciences (CAS) proposed a novel model optimization algorithm named External Calibration-Assisted Screening (ECA). This breakthrough significantly enhances the prediction robustness of Near-Infrared Spectroscopy (NIRS) quantitative models.
The findings were published in Analytica Chimica Acta.
Near-infrared spectroscopy, as a highly promising non-destructive analytical method, relies heavily on the development level of its calibration models for prediction effectiveness. However, variations in measurement conditions often cause substantial prediction deviations. Consequently, mature NIRS models require strong robustness against environmental interference.
In this study, the team proposed a paradigm shift toward robustness-oriented optimization (rather than mere accuracy) for NIRS models and introduced ECA as its implementation pathway.
The ECA method rapidly adapts initial models to new detection environments by calibrating them with externally collected samples under novel measurement conditions. The researchers innovatively integrating cross-validation results with external calibration results to establish a new robustness evaluation metric, PrRMSE, to identify the optimally robust model through multi-parameter modeling combination screening.
To further boost performance, the researchers combined ECA with an established algorithm CARS (Competitive Adaptive Reweighted Sampling). This integration led to a new optimization framework: ECCARS.
The ECCARS framework was validated using one laboratory-measured rice flour dataset and two public corn datasets. The results were impressive: compared with traditional CARS methods, ECCARS-selected models achieved a 12.15% to 725% reduction in calibration errors, and a 27.63% to 482% reduction in validation errors under varying conditions.
These results confirm that ECCARS dramatically improves the robustness and reliability of NIRS models, paving the way for more stable and accurate real-world applications.
Schematic diagram of using ECA method for model optimization. (Image by XU Zhuopin)
