應用人工智慧於早產兒視網膜病變輔助鑑別診斷之研究

台灣地區每年約有210,000名新生兒，估算每年約有10,500至21,000個家庭會面臨早產兒的出生。由於小兒科新生兒醫學的進步，使得早產兒的救護技術不斷地進步，早產兒的存活率越來越高，相對地也帶來許多疾病發生，早產兒存活率增加，導致罹患早產兒視網膜病變的危險性也增加。醫學中鮮少有針對早產兒視網膜病變做有效篩檢方式或診斷的工具。本研究的目的乃在藉由探勘技術中決策樹的分析與類神經網路的運用來建構早產兒視網膜病變之輔助鑑別診斷模式，提供醫師未來進行醫療服務時，於早期評估及判斷早產兒視網膜病變階段，減少因為人為判斷失誤所造成不必要的醫療浪費，同時可擬定早期防範的對策，降低發生的可能性。比較兩種模型後，決策樹之平均準確率為86.4%優於類神經網路的84.41%。由決策樹模型所得到的規則經與醫師確認後在臨床上是有診斷的有效性並與文獻吻合，研究結果顯示發現妊娠週數、出生體重、氧氣使用天數、呼吸器的使用、和氧氣濃度等變數是主要的危險因子。本研究建構之模型對於醫師之臨床診斷將有實質之助益。

關鍵字

早產兒視網膜病變；資料探勘；倒傳遞類神經網路；決策樹

並列摘要

There are 210,000 premature infants in Taiwan per year and it estimates that about 10,500 to 21,000 family would confront with the birth of premature infant. Due to the progress of Neonatology, the care technology of premature baby is making progress and the survival rate of premature baby is increasing. However, it also triggers more diseases comparatively. The more survival rate of premature infant would leads to the higher peril rate of meeting retinopathy of prematurity(ROP). In medical science, there are few instruments especially aimed at ROP to sieve or diagnose efficiently. This study aims to use the decision tree analysis of data mining and application of neural network to construct retinopathy of prematurity auxiliary diagnose model in order to assist doctors to evaluate and determine the early days situation about ROP in the future. It could decrease the unnecessary medical squander caused by artificial misjudgment. Simultaneously, we could draw up countermeasures to take precautions in order to lower down the happening possibility. Contrast with two models, the average accuracy of decision tree is 86.4% which is superior to the 84.41% of neural network. After confirmed by doctors, the rule from decision tree is equipped with clinical diagnostic validity and corresponds with literature. The study shows that the gestational Age, birth weight, oxygen using days, use of ventilation, and oxygen concentration those variables are main dangerous factors. The model constructed is beneficial to doctors’ clinical diagnosis.