良性前列腺肥大是中老年男性常見的泌尿科疾病,通常施以經尿道前列腺切除術,即可紓解其症狀。前列腺癌為國人癌症死亡原因第七位,死亡率達2.34%,於初期施以前列腺根除術,治癒率很高。隨著人口老化,近幾年前列腺患者有快速攀升趨勢,其治療所耗費的醫療資源與費用相當可觀,如何有效評估住院日與醫療費用對於整體醫療資源規劃是值得深入探討的醫療議題。 本研究透過文獻蒐集與醫師訪談,彙整相關影響變數,並以2006至2009年健保資料庫中前列腺手術之病患為研究對象,運用決策樹、倒傳遞類神經網路、支持向量機、案例式推理等相互結合的人工智慧方法於評估患者術後住院日與醫療費用。研究結果,住院日方面,A案例中各模型之準確率與ROC曲線下面積皆有78%及0.7以上之水準;B案例中SVM和BPN結合SVM兩者模型具有較佳之評估結果;C案例中C5.0結合SVM模型之敏感度僅24%,無法有效辨識住院超過標準值的個案。醫療費用方面,A、C兩案例平均誤差費用比率約為14%優於B案例的24.46%,代表有較佳的評估表現。本研究可提供醫療人員做為臨床輔助評估住院日與醫療費用之參考,有效提升醫療服務品質與效率,避免不必要之浪費,且對醫療機構於資源配置上有實質上的助益。
The Benign Prostatic Hyperplasia (BPH), is one of the common aged male urinary tract diseases. Usually applying the radical prostatectomy can relieve of the symptoms of BPH. The prostate cancer ranks the top 7th for domestic death rates arising from cancers and its death rate is about 2.34%. In case of conducting the radical prostatectomy in the preliminary stage, the curative ratio is pretty high. With the population remains in a state of aging and the ever-increasingly rising number of prostate patients, the corresponding consumed medical treatment resources and costs are tremendous. One of the top priority medical issues shall be show to effectively evaluate the length of stay and the relevance the medical treatment costs and overall medical resource planning. It has been conducted reference document collection and in-depth interviews with the physicians for summarizing all the relevant influential variables in conjunction with regarding the prostate surgery patients in the database of Bureau of National Health Insurance from years 2006 to 2009. The portfolios of various artificial intelligence algorithm, including C5.0 Decision Tree, Back Propagation Neural Network (BPN), Support Vector Machines (SVM), Case Based Reasoning (CBR) have been applied in this research for evaluating the length of stay after surgery and total medical treatment costs. The results of length of stay, in the case A, the accuracy and area under the ROC curve of all individual models are all maintained at least 78% and a value of 0.7. In the case B, the SVM and the BPN combined with SVM models demonstrate better analysis performance. In the case C, the C5.0 combined with SVM model yields a sensitivity ratio of only 24%, which symbolizes incompetence for making effective authentication on the cases. In terms of expenditure, the average error of expenditure for case A and case C are both about 14%, much better than the 24.46% yielding from the case B. It symbolizes a better performance. This study can providing medical professionals auxiliary reference basis for evaluating length of stay and clinical diagnosis and for substantial enhancement of medical service quality and efficiency and for avoidance of unnecessary waste and assistance for medical institutions in reaching practical benefits for medical resource appropriation. Meanwhile, the research results can also provide the Bureau of National Health Insurance some reference basis for making disbursement policy improvement measures and making the national health insurance act more comprehensive.