應用人工智慧判斷新生兒患病理性黃疸輔助鑑別診斷之研究

黃疸(jaundice)在新生兒當中是很常見的症狀。一般出生一週內的新生兒有黃疸症狀大多屬於生理性黃疸，是因為間接型膽紅素過高，在這種狀態下是屬正常現象；若發現超過生理性的範圍，必須留意是否有其他的病變，並注意嬰兒離院後膚色變化。多數人將新生兒的黃疸症狀歸咎於母乳，而忽略潛在問題的原因。通常在醫院時，嬰兒的黃疸指數維持在生理性黃疸的範圍，不用複檢，即可出院。且新生兒的體重原本就會有生理性的下降，一般情況只要七至十天後就會恢復到正常的體重範圍，但若發現嬰兒體重持續下降，身體膚色有變黃的趨勢，即應就醫觀察。本研究以判斷是否罹患黃疸症狀以及病理性黃疸新生兒是否加強照光兩種研究為目標，期望運用人工智慧的研究方法，以雲林地區某醫療機構罹患病理性黃疸之新生兒資料庫為對象，透過倒傳遞類神經網路模式、C5.0決策樹分析與類神經網路結合決策樹，決定各項變數因子的最佳權重與規則，建構準確診斷預測新生兒罹患病理性黃疸的輔助系統，能提早發現並找出其病因，接受醫護治療。研究結果顯示，類神經網路結合決策樹在判斷罹患病理性黃疸症狀有較佳的分析成效，平均測試準確率為94.72%；而在判斷病理性黃疸新生兒是否加強照光研究上，類神經網路結合決策樹的測試平均準確率為96%，顯示類神經網路結合決策樹分析模型對新生兒病理性黃疸症狀的危險因子有較佳的規則解釋能力，根據所建構之模型，以最佳的病症規則做為未來醫護人員在臨床上診斷之輔助參考依據。而及時知道有黃疸症狀的新生兒，是否為病理性黃疸，可以早期發現治療而改善。本研究能有效的解決縮短黃疸的判斷時間，對於往後醫護人員以及新生兒家屬皆能有所助益。

關鍵字

黃疸；病理性黃疸；類神經網路；決策樹；膽紅素

並列摘要

Jaundice is a very common illness for newborns. Generally “physiological jaundice” appears within the first week of a baby’s life as a result of high levels of bilirubin in the body and is usual a normal condition that causes no concern at all. However, once the bilirubin levels exceed normal range of physiological jaundice, other symptoms of possible complications should be closely monitored and the skin color of the babies require further observation after they are discharged from the hospital. Many people blame symptoms of newborn jaundice on breast feeding but often neglect the causes of other potential problems. Usually when newborns are still in the hospital, if their jaundice index maintains within the normal range of physiological jaundice, they can be discharged without re-examinations. Furthermore, newborns tend to lose weights in the first couple of days after birth, and they usually regain their birth weights in a couple of days and return to the normal growth rate in 7 to 10 days. However, once the newborn baby is found to continue losing weights while his/her skin color becoming more yellowish, they should be brought to the hospital for a thorough examination. In this study, we started out with two objectives: to determine whether babies develop jaundice symptoms and whether phototherapy should be administered more frequently and proactively. We hope to adopt research methods in artificial intelligence, using the newborn database for babies developed physiological jaundice in one certain medical institute in Yunlin area as the subject and analyzing data through the BPN model and C5.0 decision tree and BPN combining with C5.0 decision tree to determine the best weight and rule for each variable factor and to construct a support system that accurately diagnose and predict newborns with physiological jaundice hoping to identify the causes early so medical treatments can be provided. The results indicate that the BPN combining with decision tree model has better performance in determining whether babies develop physiological jaundice or not with the average test accuracy at 94.72%; on the other hand, in determining whether phototherapy should be administered on newborns with physiological jaundice, the average accuracy of the BPN combining with decision tree model is 96%, indicating that it has better rule explanation ability on the risk factors. Based on the constructed model, the best symptom rule will be used as an important referential basis in clinical diagnosis for future medical care personnel. When jaundiced newborns are identified in time, whether the jaundice is physiological or not, they can be treated properly at an early stage and thus improve symptoms. This study can help effectively reduce the time needed to determine jaundice and will be extremely beneficial for newborn families and also helpful for the medical care staffs.