人工智慧（或深度學習）自其復興和成熟以來，改變了許多生物醫學研究，甚至在五年前（2016至今）對相關研究的發展更加迅速。我們提議將應用在西藥的最先進深度學習技術擴展至中藥領域。因此，我們開發了卷積神經網絡（CNN）將中藥處方分類為對應的疾病。經過CNN訓練後，使用了來自國家衛生局保險數據庫中的報銷數據，CNN輸出以下內容的概率（稱為AI分數）對應給定輸入中草藥治療的疾病。為了更好地理解CNN的輸出以及中藥從多種草藥中合成配方的方式，我們從數百種中藥配方裡計算每個藥草的平均比例權重（稱為重要性分數）。結果顯示，AI分數和重要性分數之間的相關性表明中藥配方裡的藥草組合公式並非簡單加法（即線性）。;Artificial intelligence (or deep learning) has accelerated and even transformed many of the biomedical research since its resurgence and maturity five years ago (around 2016). We propose to extend applications of state-of-the-art deep learning techniques to traditional Chinese medicine (TCM) from modern western medicine. In doing so, we developed a convolutional neural network (CNN) to classify TCM herbal prescriptions into their corresponding diseases. After CNN training using reimbursement data from the National Health Insurance Database, the CNN outputs probabilities (called AI scores) of the indicated diseases given an input TCM herb. To make better sense of the CNN outputs and the way TCM composes formulas from multiple herbs, we then calculated the average proportional weight (called importance score) of an individual herb in hundreds of TCM formulas found in TCM classics. The result of correlation between AI scores and importance scores indicates that the functions of TCM formulas are not simple linear addition of individual herbs.
