Prediction-bias-informed slope and bias correction (SBC) for cross-batch near-infrared quantification during herbal extract concentration

Authors:
  • KYKai Yang
    Kai Yang
  • XWXinhao Wan
    Xinhao Wan
  • YYYiqin Yang
    Yiqin Yang
  • YWYuejie Wang
    Yuejie Wang
  • WZWenting Zhu
    Wenting Zhu
  • YWYaqi Wang
    Yaqi Wang
  • ZWZhenfeng Wu
    Zhenfeng Wu
  • QTQing Tao
    Qing Tao
  • XWXuecheng Wang
    Xuecheng Wang

Abstract
Near-infrared (NIR) spectroscopy offers rapid, nondestructive analysis, but maintaining the predictive performance of quantitative models across new production batches remains a challenge for reliable quality monitoring in complex manufacturing processes. This study developed NIR quantitative models for solid content, narirutin, and hesperidin during the concentration of Jianwei Xiaoshi Tablet extracts. Batches B1–B5, prepared from raw-material batch M1, were used for model development and leave-one-batch-out cross-validation, while B6, prepared from a different raw-material batch (M2), served for independent external validation. Cross-batch predictive performance was compared across different modeling methods, data augmentation, and domain adaptation approaches. Prediction bias was then characterized, and slope and bias correction (SBC) was applied to the model outputs. Most models showed reduced predictive performance on B6, with residual predictive deviation (RPD) values below 3 for all three attributes across all models. For partial least squares regression (PLSR), the root mean square errors of prediction (RMSEPs) for solid content, narirutin, and hesperidin increased by 350%, 89%, and 90%, respectively, on B6 relative to B1–B5. After SBC, the corresponding RMSEPs decreased by 78%, 58%, and 67%. Data augmentation and domain adaptation did not consistently improve external-batch prediction, whereas SBC reduced systematic prediction bias. These findings suggest that prediction-bias-informed model correction can improve the cross-batch applicability of NIR models and inform the future development of online quality monitoring systems integrating model-status assessment with correction as needed.

This publication has 27 references indexed in Scilit: