結核病與肝臟疾病一直是全世界盛行的疾病,結核病也是全球傳染病頭號死因,在過去經驗上肝硬化患者較一般人容易感染結核病,但肝硬化和結核病之間的相關性目前研究甚少。本研究以國內某醫療機構資料庫中十八歲以上肝硬化患者為研究對象,透過相關文獻和醫師訪談,篩選出容易導致結核病的重要因子後,運用粒子群最佳化演算法、基因邏輯斯迴歸演算法、交叉熵演算法來計算各因子權重,結合倒傳遞類神經網路與支援向量機以建構預測模型與案例式評估系統,評估肝硬化患者是否伴隨結核病。研究結果顯示,雖以粒子群最佳化演算法結合案例式推理評估系統準確率最佳,經K疊驗證後,平均準確率為91.52%,ROC曲線下面積為0.917,經傅利曼檢定為無顯著差異,表示模型之間並無差異,每種模型皆可做為評估系統權重值計算;六種預測模型中,雖以粒子群最佳化演算法結合倒傳遞類神經網路最佳,K疊驗證後平均準確率為91.22%,ROC曲線下面積為0.883,經傅利曼檢定為無顯著差異,表示模型間不存在顯著差異性,因此皆可做為預測之計算。本研究結果能提供給醫療機構或臨床工作者做為輔助診斷之參考依據,達成早期發現早期治療,便能減輕病患之疾病負擔把握黃金就醫時間。
Tuberculosis, TB and liver disease have long been diseases of worldwide prevalence. TB is also the leading cause of death among the infectious diseases around the world. From past experience, patients with cirrhosis are more susceptible to TB, but studies on the correlation between cirrhosis and TB remain scarce. This study adopted patients with cirrhosis above the age of 18 from the database of an anonymous domestic medical institution as research participants. Through relevant literatures and interviews with physicians and after selecting the important factors likely to cause TB, the particle swarm optimization algorithm, PSO, genetic algorithm with logistic regression, and cross-entropy algorithm were adopted to calculate the weights of the factors. The back-propagation network, BPN and support vector machine, SVM were conjunctively used to construct predictive models and case-based reasoning systems to evaluate whether patients with cirrhosis have accompanying TB. Research results show that although the PSO combined with the case-based reasoning system has the best accuracy, after the k-fold verification, the average accuracy, ACC was 91.52% and area under the ROC curve, AURC was 0.917. The Friedman’s test shows no significant difference, thus indicating no difference exists between the models. Therefore, the models are all suitable for calculating the evaluation system weights. Among the six predictive models, although the PSO combined with the BPN are the best, after the k-fold verification, the average ACC was 91.22% and AURC was 0.833. The Friedman’s test shows no significant difference, thus indicating no difference exists between the models and that they can all be used in predictive calculations. The research results shall serve as a reference for medical institutions or clinical workers during aided diagnosis, thereby achieving “early detection and early treatment”, relieving patients of the burden of disease, and ensuring timely treatment during the golden time.