Predicting Steam Processing Degree of Prepared Radix Rehmanniae (Shudihuang) Using Machine Learning

The classic way for the prepared Radix Rehmanniae (PRR) preparation, “steaming and drying (SD) for several cycles (generally nine times, SD9),” is the golden standard method from the traditional processing theory. However, the controversy of “optimal SD cycle” still exists, and there has not been an efficient way to identify the processing degree of PRRs. The study aims to determine the optimal processing conditions to make PRR approaching the SD9 quality and establish validated models to identify the processing degree of PRR unknowns. PPR-SD1–9 samples were prepared under 1 to 9 SD cycles. A spectrophotometer and a camera were used for color and gloss assessment. The chemical changes during processing were detected by high-performance liquid chromatography and liquid chromatography coupled with mass spectroscopic (LC-MS) technique. Statistical analysis of the LC-MSE data using principal component analysis showed separation of PRRs with different processing degrees, which led to the use of random forest (RF) for model training. The changes in both appearance and chemicals were obvious before, but slight after 3–5 SD cycles. Two predictors based on the RF classifier were proven to be valid for the identification of steaming time in a wide range (0–78 hours) and with an error rate of 0 in blind verification. Ten commercial samples were then identified as 1–2 SD cycle-processed PRRs. Our results, of which processing 3 to 5 times can get PRR-SD9, are aligned with the past documented knowledge in QianJinYiFang. Our models can be good tools for quality control in PRR manufacture and supervision.